技术领域technical field
本发明涉及图像处理和字符识别等技术领域，具体涉及基于caffe与软触发下的门牌压印字符识别装置。The invention relates to the technical fields of image processing and character recognition, and in particular to a house plate embossing character recognition device based on caffe and soft trigger.
背景技术Background technique
工业上对于门牌压印字符的识别装置在众多生产线中都有着极大的需求，现有的面向工业中的压印门牌字符识别装置在图像采集以及字符处理与识别方面主要存在以下几个方面的不足，有待改进：In the industry, there is a great demand for the recognition devices of house plate embossed characters in many production lines. The existing industrial-oriented embossed house number character recognition devices mainly have the following limitations in terms of image acquisition, character processing and recognition Insufficient, room for improvement:
首先，常见的压印门牌字符识别装置在图像获取以及触发拍照方面，主要集中在采用硬件触发图像采集模块进行图像采集工作，由于硬件存在着安装条件限制因素过多，例如：场地、环境、光线等条件的限制，因此往往不能很好地适应各种生产场景，同时硬件触发在获取的图像质量方面，不进行任何选择，对于成像质量不好的图片同样采集保存，造成资源浪费和识别效率、识别率的降低；其次常用装置无法对门面和门框做出产品种类上的区分，只能完成单纯的计数与拍照功能，对于后续生产过程中所需的产品种类与序列号无法提供相应的数据源，设备适用性不高，因此近年来对于该类设备在图像获取及其触发方式上均有一定程度的优化与改进。First of all, the common embossed house number character recognition device mainly focuses on image acquisition and triggering of photographing by using hardware to trigger image acquisition modules for image acquisition. Due to hardware installation conditions, there are too many limiting factors, such as: venue, environment, light and other conditions, so it is often not well adapted to various production scenarios. At the same time, the hardware trigger does not make any selections in terms of the quality of the acquired images. The images with poor imaging quality are also collected and saved, resulting in waste of resources and recognition efficiency. The recognition rate is reduced; secondly, the commonly used devices cannot distinguish the product types between the facade and the door frame, and can only complete the simple counting and photographing functions, and cannot provide corresponding data sources for the product types and serial numbers required in the subsequent production process , the applicability of the equipment is not high, so in recent years, the image acquisition and triggering methods of this type of equipment have been optimized and improved to a certain extent.
其次，目前市面上常用的压印字符识别装置在字符识别方法上主要采用基于传统的模板特征匹配以及基于结构统计的算法，例如：以字符图像的灰度作为匹配信息，通过计算字符模板图像与待识别目标图像子区域的灰度相关系数来度量匹配度；或者利用字符本身的结构形态，如边缘、拐点、连通区域等特征进行模式识别；或者在建立字符图像的特征之后，通过有监督的方式学习特征与目标类别之间的关系，从而建立从输入到输出的函数关系。基于传统的模板匹配和特征提取的识别技术，在理想环境下的字符识别效果较好；但是随着生产环境的复杂多样化，一方面人为选取字符特征往往难以找到最合适的特征，另一方面传统方法在复杂场景下的字符识别效果不好，很难实时并且准确地识别门牌字符。近年来，随着神经网络以及机器学习的兴起和发展以及相关硬件设备条件的完善，利用神经网络和深度学习框架进行数据训练，让计算机自主学习，提取待检测数据的特征，建立识别模型从而达到自主识别的效果，成为了目前工业上对于多目标检测识别的一项热门技术。Secondly, the commonly used embossed character recognition devices currently on the market mainly use traditional template feature matching and structural statistics-based algorithms in character recognition methods. For example, using the grayscale of character images as matching information, by calculating the The gray correlation coefficient of the sub-region of the target image to be recognized is used to measure the matching degree; or the structural form of the character itself, such as edges, inflection points, connected regions, etc., are used for pattern recognition; or after the characteristics of the character image are established, supervised The way to learn the relationship between features and target categories, so as to establish the functional relationship from input to output. Based on the traditional template matching and feature extraction recognition technology, the character recognition effect is better in an ideal environment; however, with the complexity and diversification of the production environment, on the one hand, it is often difficult to find the most suitable features by artificially selecting character features; The character recognition effect of the traditional method is not good in complex scenes, and it is difficult to recognize the characters of the house number in real time and accurately. In recent years, with the rise and development of neural networks and machine learning and the improvement of related hardware equipment conditions, the use of neural networks and deep learning frameworks for data training allows computers to learn independently, extract features of the data to be detected, and establish recognition models to achieve The effect of autonomous recognition has become a popular technology for multi-target detection and recognition in the industry.
本发明针对以上问题，提出了采用基于caffe与软触发下的门牌压印字符识别装置，进行门牌压印字符的识别。首先，基于红外对射装置对门牌所在产品的种类进行区分，方便生产过程中的后续工作，进一步完善产品信息，提高了装置的适用性；其次，采用软触发方式进行图像采集，避免了由硬件触发下引起的图像清晰度不高的问题，从而提高了系统图像采集能力；最后，依托于手写字体识别深度学习框架LeNet-5，通过调用caffe框架中的LeNet-5手写字体识别网络框架，提高了门牌字符识别算法的正确率、鲁棒性和计算速度，能更好地适应目前的生产需求，在工业生产上有较强的适用性。Aiming at the above problems, the present invention proposes a house plate embossed character recognition device based on caffe and soft trigger to recognize the house plate imprinted characters. Firstly, based on the infrared cross-radiation device, the type of product where the doorplate is located is distinguished, which facilitates the follow-up work in the production process, further improves product information, and improves the applicability of the device; secondly, the soft trigger method is used for image acquisition, which avoids hardware The problem of low image clarity caused by triggering improves the image acquisition capability of the system; finally, relying on the handwritten font recognition deep learning framework LeNet-5, by calling the LeNet-5 handwritten font recognition network framework in the caffe framework, it improves It improves the accuracy, robustness and calculation speed of the house number character recognition algorithm, can better adapt to the current production needs, and has strong applicability in industrial production.
发明内容Contents of the invention
本发明首先通过摄像头采集模块、红外对射模块以及工控机等硬件设备，通过软件触发的方式，完成字符图像的采集与门牌字符所在产品种类的区分工作；其次利用漫水填充算法对字符图像区域进行处理，去除背景干扰，得到完整且单一的字符区域图像，完成对字符的定位工作；同时采用基于caffe框架中的LeNet-5手写字体识别网络框架，进行字符的训练与学习，实现了对于门牌压印字符的识别功能。本发明通过以下技术方案实现：The present invention firstly completes the collection of character images and the discrimination of the product types where the house number characters are located through hardware devices such as a camera acquisition module, an infrared contrast module, and an industrial computer, and through software triggering; Process, remove background interference, obtain a complete and single character area image, and complete the character positioning; at the same time, use the LeNet-5 handwritten font recognition network framework based on the caffe framework to conduct character training and learning, and realize the house number Recognition function of embossed characters. The present invention is realized through the following technical solutions:
基于caffe与软触发下的门牌压印字符识别装置，包括硬件采集模块与软件处理程序模块两部分。所述的硬件采集模块包括摄像头采集模块、红外对射模块、集控模块；所述的软件处理程序模块包括软件触发拍照模块、图像预处理模块、字符切分模块、caffe学习与识别模块；所述的摄像头采集模块包括摄像头1号与摄像头2号；所述的红外对射模块包括红外对射模块1号与红外对射模块2号，红外对射模块1号与红外对射模块2号又分别由对射发送端与对射接收端组成。The house plate embossing character recognition device based on caffe and soft trigger includes two parts: a hardware acquisition module and a software processing program module. Described hardware acquisition module comprises camera acquisition module, infrared shooting module, centralized control module; Described software processing program module comprises software trigger camera module, image preprocessing module, character segmentation module, caffe learning and recognition module; The camera acquisition module described above includes No. 1 camera and No. 2 camera; the infrared contrast module includes No. 1 infrared contrast module and No. 2 infrared contrast module, and No. 1 infrared contrast module and No. 2 infrared contrast module They are respectively composed of a through-beam transmitter and a through-beam receiver.
基于caffe与软触发下的门牌压印字符识别装置，首先对于硬件采集模块中所述的摄像头采集模块、红外对射模块以及集控模块的摆放位置及其连接方式说明如下：第一，摄像头采集模块主要负责采集当前字符图像，应生产线上采集区域指定位置需求，对摄像头1号、2号的摆放位置做如下说明：记摄像头1号所对应的可拍摄区域范围的极左侧与摄像头2号所对应的可拍摄区域范围的极左侧形成的交点为交点a，记摄像头1号所对应的可拍摄区域范围的极右侧与摄像头2号所对应的可拍摄区域范围的极右侧形成的交点为交点b，摄像头1号与摄像头2号在生产线中的放置位置需保证交点a与交点b均位于门牌生产线传送前进方向上，在此放置位置下，可以保证产品通过生产线传送方向上时，准确完整的位于摄像头的成像视野中心区域，此时成像较为清晰完整，字符内容被完整的获取，没有断裂缺失的情况；第二，由于压印字符位于门或门框边沿两侧，因此红外对射模块1号应位于交点a左侧，且与交点b的距离应小于一个门框的宽度；红外对射模块2号位于交点b右侧，且与交点a的距离应小于一个门框的宽度；以此确保当触发拍照时仅门面可以对其中一组红外对射模块产生遮挡；第三，所述的集控模块与摄像头采集模块通过USB数据连接线连接，传输数据，集控模块与红外对射模块则通过采集卡与串口连接器相连接，传递信号。Based on the house plate embossed character recognition device under caffe and soft trigger, firstly, the placement and connection methods of the camera acquisition module, infrared contrast module and centralized control module described in the hardware acquisition module are as follows: First, the camera The acquisition module is mainly responsible for collecting the current character image. In response to the specified location requirements of the collection area on the production line, the placement of cameras No. 1 and No. 2 is explained as follows: Note that the extreme left side of the camera No. 1 corresponding to the shooting area range and the camera The intersection point formed by the extreme left side of the photographable area corresponding to No. 2 is the intersection a, which means the extreme right side of the photographable area corresponding to camera No. 1 and the extreme right side of the photographable area corresponding to camera No. 2 The formed intersection point is intersection point b. The position of camera No. 1 and camera No. 2 in the production line needs to ensure that both intersection point a and intersection point b are located in the forward direction of the door plate production line. Under this placement position, it can ensure that the product passes through the production line in the direction of transmission. At this time, it is accurately and completely located in the central area of the imaging field of view of the camera. At this time, the imaging is relatively clear and complete, and the content of the characters is completely acquired without any breakage. Second, because the embossed characters are located on both sides of the edge of the door or door frame, the infrared The No. 1 through-radiation module should be located on the left side of the intersection point a, and the distance from the intersection point b should be less than the width of a door frame; the No. 2 infrared through-radiation module should be located on the right side of the intersection point b, and the distance from the intersection point a should be less than the width of a door frame; This ensures that only the facade can block one of the infrared shooting modules when the camera is triggered; the third, the centralized control module and the camera acquisition module are connected by a USB data cable to transmit data, and the centralized control module and the infrared shooting module The radio module is connected with the serial port connector through the acquisition card to transmit the signal.
基于caffe与软触发下的门牌压印字符识别装置，其工作原理与流程如下：The working principle and process of the house plate embossing character recognition device based on caffe and soft trigger are as follows:
首先，通过硬件采集模块采集字符图像，其中摄像头采集模块主要负责采集图像；红外对射模块判断产品种类；集控模块负责运算、信号传输与图像的处理和保存。First, the character image is collected through the hardware acquisition module, in which the camera acquisition module is mainly responsible for image acquisition; the infrared beam module judges the product type; the centralized control module is responsible for calculation, signal transmission and image processing and storage.
其次，通过软件处理程序模块中的软件触发拍照模块完成图像成像的区域选择与拍照；图像预处理模块对字符图像的处理加工，定位字符位置；字符切分模块对定位之后的训练字符、测试字符与待检测字符进行切分，形成相应的字符数据集；caffe学习与识别模块对字符数据集的训练与学习，完成最后的识别工作。Secondly, trigger the photographing module through the software in the software processing program module to complete the area selection and photographing of image imaging; the image preprocessing module processes the character image, and locates the character position; Segment with the character to be detected to form a corresponding character data set; the caffe learning and recognition module trains and learns the character data set to complete the final recognition work.
现对装置中各个模块的工作原理进行详细说明：The working principle of each module in the device is now described in detail:
当门牌所在产品经由摄像头采集模块与红外对射模块，经生产线传输时，首先通过软件处理程序模块中所述的软件触发拍照模块进行检测，判断当前采集画面中有无符合设定尺寸大小的圆形存在(需要说明的是：在此规定默认待识别压印门牌字符区域中有一压印圆形，内部压印有代表产品等级的字符，且其尺寸大小固定，可以作为对产品的筛选标准)，当检测到当前采集图像中存在唯一且清晰的符合设定尺寸大小的圆形时，说明当前采集区域有效，则通过软触发的方式对当前的字符图像进行采集拍照，同时检测红外对射模块1号与红外对射模块2号的通断状态：当红外对射模块1号与红外对射模块2号同时为导通状态时，判断当前门牌字符所在产品种类为门框；其他状态下判断当前门牌字符所在产品种类为门面，从而对产品种类进行区分，同时对当前采集的图像进行保存和记录，对于门框和门面，在判别种类之后，为了使用者方便直观的区分，在文件名命名的第三位上，将门框记为1，门面记为0，以示区分，为之后的生产过程中提供更加完整的产品数据信息。When the product where the house number is located is transmitted through the production line through the camera acquisition module and the infrared shooting module, the software described in the software processing program module first triggers the camera module to detect whether there is a circle that meets the set size in the current acquisition picture. (It should be noted that there is an embossed circle in the area of the embossed house plate characters to be recognized by default, and the characters representing the product grade are embossed inside, and its size is fixed, which can be used as a screening standard for products) , when it is detected that there is a unique and clear circle that meets the set size in the current acquisition image, it means that the current acquisition area is valid, and the current character image is acquired and photographed by soft triggering, and the infrared radiation module is detected at the same time On-off status of No. 1 and infrared inter-radiation module No. 2: when infrared inter-radiation module No. The product category where the house plate characters are located is the facade, so as to distinguish the product category, and at the same time save and record the currently collected images. For the three digits, record the door frame as 1 and the facade as 0 to show the distinction and provide more complete product data information for the subsequent production process.
本发明在图像采集阶段采用的拍照触发方式为软触发，相较于硬件触发拍照方式而言，软触发对装置所在环境以及周围的硬件条件要求不高，且无需安装硬件设备，其灵活性更强，通过对门牌字符区域有无标准圆形的检测，进一步限定了摄像头采集图像的范围，对于成像质量不好、圆形检测尺寸或清晰度不符合要求、成像区域中并没有所需待识别字符区域的图像不予采集，既缩小了采集范围，也降低了后续图像处理工作的计算量，提高了装置的处理速度与适用能力。The photographing trigger mode adopted in the image acquisition stage of the present invention is soft triggering. Compared with the hardware triggering photographing mode, soft triggering has less requirements on the environment where the device is located and the surrounding hardware conditions, and there is no need to install hardware equipment, and its flexibility is more Strong, through the detection of whether there is a standard circle in the character area of the house number, the range of image acquisition by the camera is further limited. For poor imaging quality, circle detection size or definition does not meet the requirements, and there is no required recognition in the imaging area The image in the character area is not collected, which not only reduces the collection range, but also reduces the calculation amount of subsequent image processing work, and improves the processing speed and applicability of the device.
一、对于采集得到的字符图像，首先经软件处理程序模块中所述的图像预处理模块，完成对字符图像的处理加工以及字符位置的准确定位，得到不含背景干扰的纯字符图像。该模块的具体处理步骤流程如下：1. For the collected character images, first, through the image preprocessing module described in the software processing program module, the processing of the character images and the accurate positioning of the character positions are completed to obtain a pure character image without background interference. The specific processing steps of this module are as follows:
步骤一：对于采集得到的字符图像首先进行图像灰度化，在减少图像处理数据量的同时保证图像中的基本信息不丢失；其次进行低通高斯滤波与中值滤波，对图像中的噪声点进行初步去除，进一步提高图像质量；之后利用sobel算子对图像X方向上进行梯度运算，(此处只对图像X方向上进行梯度运算原因在于：此处进行边缘检测的目的在于初步获取字符的位置信息，而根据先验知识可知字符位置集中压印在门牌的竖直方向上，因此只需要对图像进行X方向的梯度运算即可)；再进行图像二值化，将字符图像转化为黑白二值图像，去除部分背景，获得字符区域的初步轮廓；最后对二值化之后的图像进行形态学运算，消除字符中的细小断裂区域以及细小的噪声点，将短小像素点进行连接合并，从而进一步去除非字符区域的干扰。Step 1: For the collected character image, the image is grayed first, and the basic information in the image is not lost while reducing the amount of image processing data; secondly, low-pass Gaussian filtering and median filtering are performed to filter the noise points in the image Perform preliminary removal to further improve the image quality; then use the sobel operator to perform gradient calculations on the X direction of the image (here only perform gradient calculations on the X direction of the image because: the purpose of edge detection here is to initially obtain the character's position information, and according to prior knowledge, it can be known that the character position is concentrated on the vertical direction of the house number, so only the gradient calculation in the X direction is required for the image); then image binarization is performed to convert the character image into black and white Binary image, part of the background is removed, and the preliminary outline of the character area is obtained; finally, the morphological operation is performed on the image after binarization, the small broken areas and small noise points in the character are eliminated, and the short pixels are connected and merged, so that Further remove interference from non-character areas.
步骤二：对于经过步骤一之后的字符图像，进行门牌字符轮廓的检测与定位：首先通过边缘轮廓搜索函数，大致搜索字符图像的轮廓区域，返回轮廓边界的位置坐标矩阵contours，进而求取其最小外接矩形；其次对获得的外接矩形利用漫水填充算法，去除字符图像背景中的非字符区域，只保留门牌字符区域。Step 2: For the character image after step 1, detect and locate the contour of the house number character: firstly, use the edge contour search function to roughly search the contour area of the character image, return the position coordinate matrix contours of the contour boundary, and then find its minimum The circumscribed rectangle; secondly, use the flood filling algorithm on the obtained circumscribed rectangle to remove the non-character area in the background of the character image, and only keep the character area of the house number.
步骤三：对于经过步骤二之后的字符图像，利用边缘检测Canny算子对字符图像做边缘检测，获得字符图像的大致边缘轮廓；其次，通过霍夫变换函数检测图中满足特定要求的直线段。在此需要说明的是，根据先验知识，字符所在区域大致位于整幅图像的1/3至2/3区域，因此首先确定检测区域为整幅图像的1/3至2/3区域，对于该区域中依次绘制出利用霍夫变换函数检测到的每条线段，对于长度不满足基本像素点长度要求的线段，进行筛选。Step 3: For the character image after step 2, use the edge detection Canny operator to perform edge detection on the character image to obtain the approximate edge contour of the character image; secondly, use the Hough transform function to detect the straight line segment that meets the specific requirements in the graph. What needs to be explained here is that, according to prior knowledge, the area where the character is located is roughly located in the 1/3 to 2/3 area of the entire image, so first determine that the detection area is 1/3 to 2/3 area of the entire image, for In this area, each line segment detected by the Hough transform function is drawn sequentially, and the line segment whose length does not meet the length requirement of the basic pixel point is screened.
经过霍夫变换检测直线之后，图像中只保留下长度大于一定像素个数且两条直线之间间隔大于设定阈值(此处阈值可以根据经验值做调整，本算法中设置为20个像素)的直线段，并对这些最终检测得到的直线线段按照坐标位置进行排序，按照直线所在位置，顺序返回每条直线的起点与终点的X、Y坐标，根据起点终点坐标，最终可以确定字符所在区域的边界线，记做图像dst，最终可以获得门牌字符的确定位置，由此完成字符轮廓的检测与定位工作。After the straight line is detected by the Hough transform, only the length greater than a certain number of pixels and the interval between the two straight lines are greater than the set threshold are retained in the image (the threshold here can be adjusted according to the empirical value, which is set to 20 pixels in this algorithm) The straight line segments of these final detections are sorted according to the coordinate positions, and the X and Y coordinates of the starting point and the end point of each straight line are returned in order according to the position of the straight line. According to the coordinates of the starting point and the end point, the area where the character is located can be finally determined The boundary line of , recorded as the image dst, can finally obtain the definite position of the house plate character, thereby completing the detection and positioning of the character outline.
步骤四：对于经过步骤三之后的字符图像dst，首先计算字符图像上下边界两直线的斜率，分别记为K1和K2，对于K1和K2，求其均值，记为K，进而计算出旋转角度；其次，以字符图像区域的中心点坐标为旋转中心，将字符图像dst进行旋转得到旋转之后的图像dst2，目的在于将字符图像旋转成水平方向，方便后续的字符切分和识别。Step 4: For the character image dst after step 3, first calculate the slope of the two straight lines on the upper and lower boundaries of the character image, which are respectively denoted as K1 and K2, and for K1 and K2, find their mean value, denoted as K, and then calculate the rotation angle; Secondly, the character image dst is rotated to obtain the rotated image dst2 with the coordinates of the center point of the character image area as the center of rotation. The purpose is to rotate the character image to a horizontal direction to facilitate subsequent character segmentation and recognition.
经以上处理步骤之后，可获得完整且没有背景干扰因素的纯字符图像区域，且字符压印顺序符合常见的文字顺序，方向为水平方向排列，为之后的字符切分模块提供了图像数据。After the above processing steps, a complete pure character image area without background interference factors can be obtained, and the character embossing order conforms to the common character order, and the direction is arranged in the horizontal direction, which provides image data for the subsequent character segmentation module.
二、对通过上述步骤得到的字符图像，经过字符切分模块处理，得到切分后的单个字符图像，并且标准化为大小一致的切分结果图，结果图将在后续识别模块中作为训练数据源。该模块的具体处理步骤流程如下：2. The character image obtained through the above steps is processed by the character segmentation module to obtain a single character image after segmentation, and is standardized into a segmentation result image of the same size. The result image will be used as a training data source in the subsequent recognition module . The specific processing steps of this module are as follows:
步骤一：首先，对字符图像dst2，以图像的左上角作为像素搜索的起点，对于整张图片，从上至下逐行扫描判断图像中每一点的像素值，主要处理流程如下：Step 1: First, for the character image dst2, use the upper left corner of the image as the starting point for pixel search, and for the entire image, scan from top to bottom line by line to determine the pixel value of each point in the image. The main processing flow is as follows:
1.判断该点像素值是否为0，若该点像素值为0(表示该点为黑色)，则继续扫描下一点像素值；若该点像素值为1(表示该点为白色)，则进行判断2。1. Determine whether the pixel value of the point is 0, if the pixel value of the point is 0 (indicating that the point is black), then continue to scan the pixel value of the next point; if the pixel value of the point is 1 (indicating that the point is white), then Make a judgment 2.
2.判断该点纵坐标是否处于图片整体宽度的1/3到2/3之间，对于纵坐标不满足要求的则返回判断条件1，进行下一个点的判断；若满足要求，则进行判断3。2. Determine whether the vertical coordinate of the point is between 1/3 and 2/3 of the overall width of the picture. If the vertical coordinate does not meet the requirements, return to judgment condition 1 to judge the next point; if it meets the requirements, proceed to judgment 3.
3.取该点纵坐标为基准，向左向右各30个像素范围内的点，统计该范围内点的像素值为1的像素点个数，判断满足要求的像素点个数是否大于该区域所有像素点个数的80％，若不满足要求，则返回判断条件1，进行下一个点的判断；若满足要求则返回该点的坐标值，记为点P1(X1，Y1)。3. Take the ordinate of the point as the reference, count the points within the range of 30 pixels from the left to the right, count the number of pixels with a pixel value of 1 in the range, and judge whether the number of pixels that meet the requirements is greater than this If 80% of all pixel points in the area do not meet the requirements, return to judgment condition 1 to judge the next point; if the requirements are met, return the coordinate value of this point, which is recorded as point P1 (X1, Y1).
步骤二：其次，对字符图像dst2，以图像的左下角作为像素搜索的起点，对于整张图片，从下至上逐行扫描判断图像中每一点的像素值，其主要的处理流程同步骤一中的处理流程，由此得到的点坐标值记为P2(X2，Y2)。Step 2: Secondly, for the character image dst2, use the lower left corner of the image as the starting point for pixel search, and for the entire image, scan and judge the pixel value of each point in the image line by line from bottom to top, and its main processing flow is the same as in step 1 The processing flow of the point coordinate value thus obtained is denoted as P2(X2, Y2).
步骤三：对字符图像dst2，以图像的左上角作为像素搜索的起点，对于整张图片，从左至右逐列扫描判断图像中每一点的像素值，其主要的处理流程同步骤一中的处理流程，由此得到的点坐标值记为P3(X3，Y3)。Step 3: For the character image dst2, use the upper left corner of the image as the starting point for pixel search. For the entire image, scan from left to right to determine the pixel value of each point in the image. The main processing flow is the same as that in step 1. Processing flow, the obtained point coordinate value is denoted as P3(X3, Y3).
步骤四：对字符图像dst2，以图像的左上角作为像素搜索的起点，对于整张图片，从右至左逐列扫描判断图像中每一点的像素值，其主要的处理流程同步骤一中的处理流程，由此得到的点坐标值记为P4(X4，Y4)。Step 4: For the character image dst2, the upper left corner of the image is used as the starting point for pixel search. For the entire image, scan from right to left column by column to determine the pixel value of each point in the image. The main processing flow is the same as in step 1. Processing flow, the obtained point coordinate value is denoted as P4(X4, Y4).
步骤五：根据步骤一至步骤四得到的四个点的坐标值，根据以下公式求出字符的具体位置，框定矩形框，进一步缩小字符位置所在范围：Step 5: According to the coordinate values of the four points obtained in steps 1 to 4, calculate the specific position of the character according to the following formula, frame a rectangular frame, and further narrow the range of the character position:
定义height与width分别为矩形的宽和长，记：Define height and width as the width and length of the rectangle respectively, remember:
height＝|Y2-Y1|height＝|Y2-Y1|
width＝|X4-X3|width=|X4-X3|
这样在原图字符图像上和二值化之后的图像dst2上均框定一矩形框，认为该矩形框即为字符的精确外边框。In this way, a rectangular frame is framed on both the original character image and the binarized image dst2, and the rectangular frame is considered to be the precise outer frame of the character.
步骤六：对确定精确外边框之后的字符图像，需要进一步获得单个的字符切分结果图，主要处理流程如下：Step 6: For the character image after the precise outer border is determined, it is necessary to further obtain a single character segmentation result map. The main processing flow is as follows:
1.对于确定精确外边框之后的字符图像，首先根据先验知识，分割汉字字符：按照W3，H3的长宽数据(W3，H3根据先验知识可知是压印字符模型中汉字字符的宽度和长度)，在字符图像上以W3为标准，先确定第一个字符，即汉字字符的大致右边界；再以汉字字符的大致右边界为起点，从右向左逐列扫描字符图像，遍历每一列字符图像中的像素值，当且仅当该点像素值为1(表示该点为白色)且取该点横坐标为基准，向上向下各30个像素范围内的点，统计该范围内点的像素值为1的像素点个数，满足要求的像素点个数大于该区域所有像素点个数的80％时，记录当前点的横坐标，以该横坐标为轴，所在直线即为汉字字符的精确右边界。由此即可得到第一个字符，即汉字字符的精确位置。1. For the character image after the precise outer frame is determined, firstly, segment the Chinese characters according to the prior knowledge: according to the length and width data of W3, H3 (W3, H3 can be known according to the prior knowledge to be the width and width of the Chinese characters in the embossed character model Length), using W3 as the standard on the character image, first determine the first character, that is, the approximate right boundary of the Chinese character; then start from the approximate right boundary of the Chinese character, scan the character image column by column from right to left, and traverse each The pixel value in a column of character images, if and only if the pixel value of the point is 1 (indicating that the point is white) and the abscissa of the point is taken as the reference, and the points within the range of 30 pixels up and down are counted. The number of pixels with a pixel value of 1, and when the number of pixels that meet the requirements is greater than 80% of all the pixels in the area, record the abscissa of the current point. Taking the abscissa as the axis, the straight line is The exact right boundary of a Kanji character. From this, the first character, that is, the precise position of the Chinese character can be obtained.
2.对于之后的字母以及数字随机构成的6个字符，同样根据W1，H1、W2，H2(假设数字0-9模板的字符大小为W1，H1；字母A-Z模板大小为W2，H2)的大小宽度，先确定字符的大致右边界，再从大致右边界自右向左逐列扫描字符图像，遍历每一列字符图像中的像素值，进而确定字符的精确右边界，由此依次得到每一个字符的精确位置。2. For the following 6 characters composed of random letters and numbers, also according to the size of W1, H1, W2, H2 (assuming that the character size of the number 0-9 template is W1, H1; the size of the letter A-Z template is W2, H2) Width, first determine the approximate right boundary of the character, then scan the character image column by column from right to left from the approximate right boundary, traverse the pixel values in each column of character image, and then determine the precise right boundary of the character, and then get each character in turn precise location.
3.对于精确分割之后的单个字符，最后将其大小归一化为28*28的图片，因为后续网络模型要求其输入数据大小为28*28，将经过大小归一化之后的单个字符记为切分结果图。切分结果图按照对应的单个字符的内容保存在对应的文件夹下，字符内容与文件夹名一一对应。3. For a single character after accurate segmentation, its size is finally normalized to a 28*28 picture, because the subsequent network model requires its input data size to be 28*28, and the single character after size normalization is recorded as Segmentation result graph. The segmentation result graph is saved in the corresponding folder according to the content of the corresponding single character, and the character content corresponds to the folder name one by one.
在此需要说明的是：What needs to be explained here is:
1.该装置对应的门牌压印字符，字符均由模板压印而成，字符主要有数字、字母以及“甲乙丙丁”四类汉字构成，相同类别的字符压印模板大小统一，即所有的数字模板大小、所有的字母模板大小以及甲乙丙丁四种汉字的模板大小相同，但是这三种字符之间大小不同。1. The house plate embossing characters corresponding to the device are all embossed by templates. The characters are mainly composed of numbers, letters and "A, B, C, D" four types of Chinese characters. The embossing templates of the same type of characters have a uniform size, that is, all numbers The template size, the template size of all letters, and the template sizes of the four Chinese characters A, B, C and D are the same, but the sizes of these three characters are different.
2.由于压印字符中，字母I与数字1、字母O与数字0难以区分，因此，该装置默认压印字符中没有字母I与O。2. Since it is difficult to distinguish between the letter I and the number 1, and the letter O and the number 0 in the embossed characters, the device does not have the letters I and O in the embossed characters by default.
3.并且根据先验知识，假设数字0-9模板的字符大小为W1，H1；字母A-Z模板大小为W2，H2；汉字甲乙丙丁模板大小为W3，H3。汉字字符的尺寸大小最大，而字母和数字字符的尺寸大小相同，均略小于汉字字符。3. And according to prior knowledge, it is assumed that the character size of the template for numbers 0-9 is W1, H1; the size of the template for letters A-Z is W2, H2; the size of the template for Chinese characters A, B, C, D is W3, H3. Kanji characters are the largest in size, while alphabetic and numeric characters are the same size, slightly smaller than Kanji characters.
4.该门牌压印字符从左往右的排列规则是：从左至右第一个字符为汉字，后续字符为数字和字母随机构成，共计7个字符。4. The rules for the arrangement of the imprinted characters of the house number from left to right are: the first character from left to right is a Chinese character, and the subsequent characters are randomly composed of numbers and letters, a total of 7 characters.
5.切分结果图在后续识别模块中作为训练数据源，训练网络的称为训练切分结果图；在后续识别模块中作为测试数据源，测试网络优劣的称为测试切分结果图；在后续识别模块中作为待识别字符数据集的称为待识别切分结果图。5. The segmentation result graph is used as a training data source in the subsequent recognition module, and the training network is called a training segmentation result graph; in the subsequent recognition module, it is used as a test data source, and the test network is called a test segmentation result graph; In the subsequent recognition module, the character data set to be recognized is called the segmentation result graph to be recognized.
6.根据实际采集得到的图片数量，将其中的6万张字符图作为训练切分结果图的数据源，另采用1万张采集到的字符图作为测试切分结果图的数据源，需要注意的是此处的6万张训练样本源与1万张测试样本源互相独立不重合。6. According to the actual number of pictures collected, 60,000 character images are used as the data source of the training segmentation result image, and another 10,000 collected character images are used as the data source of the test segmentation result image. Note that The most important thing is that the source of 60,000 training samples and the source of 10,000 testing samples are independent and do not overlap.
7.由于门牌压印字符除去首字符固定为汉字之外，后续6个字符的种类和内容均为随机产生的，因此不能保证不同种类下数字和字母的切分结果图数量均匀分布，对于个别种类下的字符切分结果图如果数量过少，可以采用基本的图像处理方法(例如：旋转、去噪声、模糊、形态学操作等等)生成，进而补全切分结果图。7. Since the imprinted characters of the house number except the first character is fixed as a Chinese character, the type and content of the subsequent 6 characters are randomly generated, so it cannot be guaranteed that the numbers and letters of different types are evenly distributed. For individual If the number of character segmentation result maps under the category is too small, basic image processing methods (such as: rotation, denoising, blurring, morphological operations, etc.) can be used to generate, and then the segmentation result map is completed.
根据以上所述的六个步骤，经过字符切分模块之后最终即可得切分后的单个字符图像结果，并且按照字符内容一一对应保存在相对应的文件夹下，形成相应的切分结果图，为之后的字符识别做准备。According to the above six steps, after the character segmentation module, the result of the segmented single character image can be finally obtained, and stored in the corresponding folder according to the character content one by one to form the corresponding segmentation result Figure, in preparation for subsequent character recognition.
四.对经过字符切分模块处理得到的切分结果图，经caffe框架中的手写字体识别网络，进行网络模型的学习，并进而测试当前网络模型的好坏，最后将待识别字符与训练学习得到的网络相结合，完成最后的字符内容的识别工作。4. For the segmentation result graph processed by the character segmentation module, learn the network model through the handwritten font recognition network in the caffe framework, and then test the quality of the current network model, and finally combine the characters to be recognized with the training learning The obtained networks are combined to complete the final character content recognition work.
caffe学习与识别模块又包括学习模块与识别模块，其中学习模块具体处理流程如下(本文中所涉及文件路径名默认从caffe根目录开始)：The caffe learning and recognition module also includes a learning module and a recognition module. The specific processing flow of the learning module is as follows (the file path names involved in this article start from the root directory of caffe by default):
步骤一：在该装置进行字符识别之前，需要对网络模型进行学习，要有大量的样本输入，因此对于训练切分结果图先进行人工分类，将对应字符分别保存到对应命名的文件夹下，形成训练样本集。Step 1: Before the device performs character recognition, the network model needs to be learned, and a large number of sample inputs are required. Therefore, manual classification is performed on the training segmentation result graph, and the corresponding characters are saved in correspondingly named folders. Form a training sample set.
步骤二：对于测试切分结果图进行人工分类，将对应字符分别保存到对应命名的文件夹下，形成测试样本集。Step 2: Manually classify the test segmentation results, and save the corresponding characters in correspondingly named folders to form a test sample set.
步骤三：将训练样本集中的图片按照顺序打上标签，打乱顺序，形成标签文件，并记录标签文件所在路径，对于测试样本集中的图片做相同的操作，生成相应的标签文件。Step 3: Label the pictures in the training sample set in order, shuffle the order, form a label file, and record the path of the label file, do the same operation for the pictures in the test sample set, and generate the corresponding label file.
步骤四：图片格式转换。由于LeNet-5网络中要求输入的数据格式是ldb或者lmdb格式，因此，需要进行图片格式的转换。格式转换的具体步骤如下：Step 4: Image format conversion. Since the input data format required in the LeNet-5 network is ldb or lmdb format, it is necessary to convert the image format. The specific steps of format conversion are as follows:
1.在caffe根目录路径下自带有图片格式转换的cpp文件和对应的exe工具。1. In the caffe root directory path, it comes with a cpp file for image format conversion and the corresponding exe tool.
2.在caffe根目录路径下，创建create_mnist.bat文件(windows批量操作文件)并将其中的内容做修改。2. Under the caffe root directory path, create a create_mnist.bat file (windows batch operation file) and modify its contents.
3.修改完毕之后，运行bat文件，处理完毕之后将会在设置的需要保存的对应路径下，生成相应的测试集与训练集文件夹。3. After the modification is completed, run the bat file. After the processing is completed, the corresponding test set and training set folders will be generated under the corresponding path that needs to be saved.
步骤五：对应的修改caffe包库下LeNet-5网络中的参数与文件所在路径。步骤六：运行bat文件，生成lenet_iter_10000.caffemodel模型。Step 5: Correspondingly modify the parameters and file path in the LeNet-5 network under the caffe package library. Step 6: Run the bat file to generate the lenet_iter_10000.caffemodel model.
步骤七：测试网络模型优劣。为了测试步骤六中所得的网络模型对于其他非训练样本集中的图片识别效果如何，利用测试样本集中的图片对网络进行测试。需要注意的是，6万张训练样本与1万张测试样本是互相独立不重合的，因此，可以测试出该网络对不同数据的优劣性能，此步骤也是变相的增大了训练样本的数量，对于网络有更好的训练效果。根据实验结果可知该模型对于测试集中的字符图像数据准确率很高且不会发生过拟合以及陷入局部最优的循环中。Step 7: Test the pros and cons of the network model. In order to test how well the network model obtained in step 6 can recognize pictures in other non-training sample sets, use the pictures in the test sample set to test the network. It should be noted that the 60,000 training samples and the 10,000 test samples are independent of each other and do not overlap. Therefore, the performance of the network on different data can be tested. This step also increases the number of training samples in disguise. , which has a better training effect on the network. According to the experimental results, it can be seen that the model has a high accuracy rate for the character image data in the test set and will not overfit and fall into a local optimal cycle.
经过以上步骤一至步骤七的处理之后，即可得到训练优化之后的模型文件caffemodel，后续将利用它，对待识别字符图像进行识别。After the above steps 1 to 7, the model file caffemodel after training and optimization can be obtained, which will be used later to recognize the character image to be recognized.
五.caffe学习与识别模块中的识别模块的具体处理流程如下：步骤一：在装置运行过程中，基于硬件采集设备模块所采集到的待识别的字符图像，同样经过上述流程得到待识别切分结果图。5. The specific processing flow of the recognition module in the caffe learning and recognition module is as follows: Step 1: During the operation of the device, based on the character image to be recognized collected by the hardware acquisition device module, the segmentation to be recognized is also obtained through the above process Result graph.
步骤二：将步骤一中所得的待识别切分结果图按照caffe学习与识别模块中的学习模块中步骤四进行处理，得到格式转换之后的数据集，记为待识别字符数据集。Step 2: Process the segmentation result image obtained in step 1 according to step 4 in the learning module of the caffe learning and recognition module, and obtain the data set after format conversion, which is recorded as the character data set to be recognized.
步骤三：调用caffe学习与识别模块中学习模块所得的模型文件caffemodel，将待识别字符数据放入模型中进行识别，最终输出与模型库中对比相似度概率值最高的对应字符，即当前字符的识别结果，再进行组合输出，即为最终的门牌压印字符识别结果，由此完成最终的字符识别。Step 3: Call the model file caffemodel obtained from the learning module in the caffe learning and recognition module, put the character data to be recognized into the model for recognition, and finally output the corresponding character with the highest similarity probability value compared with the model library, that is, the current character The recognition results are then combined and output, which is the final door plate imprinted character recognition result, thus completing the final character recognition.
本发明的优点：Advantages of the present invention:
1、本发明采用caffe深度学习网络作为计算工具，依托调用LeNet-5卷积神经网络进行卷积运算，采用训练数据集与测试数据集分开训练学习的方式。训练数据集训练生成模型文件，测试数据集对训练生成的模型进行测试，所采用的测试数据集与训练数据集独立分开，使得模型在训练过程中见过更多有差异的数据，从而进一步提高了模型的鲁棒性以及泛化能力，针对实际数据也有较好的鲁棒性。1. The present invention adopts the caffe deep learning network as a computing tool, relies on calling the LeNet-5 convolutional neural network to perform convolution operations, and adopts a training data set and a test data set to train and learn separately. The training data set is trained to generate model files, and the test data set is used to test the model generated by training. The test data set used is independently separated from the training data set, so that the model has seen more different data during the training process, thereby further improving It not only improves the robustness and generalization ability of the model, but also has better robustness against actual data.
2、本发明训练数据速度快，算法自动化程度高，使用方便，效率高。2. The training data speed of the present invention is fast, the algorithm automation degree is high, the use is convenient, and the efficiency is high.
3、本发明采用软触发方式触发摄像机拍照采集图像，避免了在硬件触发方式下引起的图像清晰度不高的问题，对于成像质量较差的图片进行过滤筛选，从而提高了系统图像采集能力。3. The present invention adopts a soft trigger mode to trigger the camera to take pictures and collect images, avoiding the problem of low image definition caused by the hardware trigger mode, and filters and screens pictures with poor imaging quality, thereby improving the image acquisition capability of the system.
4、本发明将手写字体卷积网络架构LeNet-5应用于安全门生产过程中的字符识别需求中，对LeNet-5网络进行了推广应用。4. The present invention applies the handwritten font convolutional network architecture LeNet-5 to the character recognition requirements in the production process of security doors, and promotes the application of the LeNet-5 network.
5、本发明借助红外对射对门牌所在产品的种类进行区分标记，对应图像也进行区别保存，方便生产过程中的后续工作，进一步完善产品信息，进一步提高了装置的适用性。5. The present invention distinguishes and marks the types of products where the house number is located by means of infrared radiation, and also distinguishes and saves the corresponding images, which facilitates follow-up work in the production process, further improves product information, and further improves the applicability of the device.
6、本发明在定位字符所在位置过程中，采用了一种漫水填充算法来获得字符的位置信息，准确完整的消除了图片的背景区域，从而提高装置的字符提取能力，使其适用于更多更复杂的背景环境下。6. In the process of locating the position of the character, the present invention adopts a flood filling algorithm to obtain the position information of the character, which accurately and completely eliminates the background area of the picture, thereby improving the character extraction ability of the device and making it suitable for more in more complex contexts.
附图说明Description of drawings
图1装置整体结构示意图；The schematic diagram of the overall structure of the device in Fig. 1;
图2装置硬件采集设备模块安装位置示意图；Fig. 2 is a schematic diagram of the installation position of the device hardware acquisition equipment module;
图3装置摄像头采集模块成像示意图；Figure 3 is a schematic diagram of imaging of the camera acquisition module of the device;
图4装置软件触发拍照模块流程图；Fig. 4 device software triggers the photographing module flow chart;
图5装置图像预处理模块流程图；Figure 5 device image preprocessing module flow chart;
图6装置字符切分模块流程图；Fig. 6 device character segmentation module flowchart;
图7装置门牌外观结构示意图；Figure 7 is a schematic diagram of the appearance and structure of the house plate of the device;
具体实施方案specific implementation plan
下面结合附图和实施案例对本发明专利作进一步的说明，但并不作为对本发明专利限制的依据。The patent of the present invention will be further described below in conjunction with the accompanying drawings and examples of implementation, but it is not used as a basis for limiting the patent of the present invention.
如图1装置整体结构示意图所示，本装置由硬件采集设备模块1与软件处理程序模块2两部分组成，其中硬件采集设备模块1包括摄像头采集模块3、红外对射模块4、集控模块5；软件处理程序模块2包括软件触发拍照模块6、图像预处理模块7、字符切分模块8、caffe学习与识别模块9；摄像头采集模块3又包括摄像头1号10与摄像头2号11；红外对射模块4包括红外对射模块1号12与红外对射模块2号13，红外对射模块1号12与红外对射模块2号13又分别由对射发送端与对射接收端组成，集控模块5与摄像头采集模块3通过USB数据连接线连接，传输数据，集控模块5与红外对射模块4则通过采集卡与串口连接器相连接，传递信号。As shown in the schematic diagram of the overall structure of the device in Figure 1, the device is composed of a hardware acquisition device module 1 and a software processing program module 2, wherein the hardware acquisition device module 1 includes a camera acquisition module 3, an infrared radiation module 4, and a centralized control module 5 The software processing program module 2 includes a software trigger camera module 6, an image preprocessing module 7, a character segmentation module 8, and a caffe learning and recognition module 9; the camera acquisition module 3 includes a camera 1 No. 10 and a camera 2 No. 11; The shooting module 4 includes infrared shooting module 1 No. 12 and infrared shooting module 2 No. 13, and infrared shooting module 1 No. 12 and infrared shooting module 2 No. 13 are respectively composed of a shooting end and a receiving end. The control module 5 is connected with the camera acquisition module 3 through a USB data cable to transmit data, and the centralized control module 5 and the infrared shooting module 4 are connected with the serial port connector through the acquisition card to transmit signals.
基于软触发下的门牌16压印字符识别装置首先通过基于摄像头采集设备3、红外对射模块4以及集控模块5(工控机)等硬件设备完成字符图像的采集与门牌16字符所在产品种类的区分工作；其次对字符图像进行预处理，利用漫水填充算法对字符图像区域进行处理，去除背景像素的干扰，得到完整且单一的字符图像，完成对压印字符的定位工作；同时采用基于caffe深度学习与LeNet-5手写字体识别网络框架，进行网络的训练与学习，最终实现了对于门牌16压印字符的识别功能。The house plate 16 embossed character recognition device based on the soft trigger first completes the collection of character images and the product category of the house plate 16 characters through hardware devices such as the camera acquisition device 3, the infrared shooting module 4, and the centralized control module 5 (industrial computer). Differentiate the work; secondly, preprocess the character image, use the flood filling algorithm to process the character image area, remove the interference of background pixels, obtain a complete and single character image, and complete the positioning of the embossed character; at the same time, use the caffe-based Deep learning and LeNet-5 handwritten font recognition network framework, network training and learning, and finally realized the recognition function of 16 embossed characters on the house number.
该装置运行过程中，如图2装置硬件采集设备模块安装位置示意图、图3装置摄像头采集模块成像示意图、图4装置软件触发拍照模块流程图以及图7装置门牌外观结构示意图所示，根据摄像头采集模块3与红外对射模块4之间的硬件摆放位置，当压印门牌16字符所在产品通过生产线进行传输，通过摄像头采集模块3，进入摄像头1号10与摄像头2号11的成像区域时，需要保证当前门牌16压印字符在摄像头1号10与摄像头2号11的成像区域中央：首先触发红外对射模块4中对射管的状态发生改变，其次，通过软件处理程序模块2中所述的软件触发拍照模块6进行检测，判断当前采集画面中有无符合设定尺寸大小的圆形存在，当检测到当前采集画面中存在唯一、清晰且符合设定尺寸大小的圆形时，说明当前采集区域有效，则通过软触发的方式由集控模块5对当前的字符图像进行采集拍照；同时检测红外对射模块1号12与红外对射模块2号13的通断状态：当红外对射模块1号12与红外对射模块2号13同时为导通状态时，判断当前门牌16字符所在产品种类为门框14；其他状态下判断当前门牌16字符所在产品种类为门面15，从而对产品种类进行区分，并对当前采集的图像进行保存和记录，对于门框14和门面15，在判别种类之后，为了使用者方便直观的区分，在文件名命名的第三位上，将门框14记为1，门面15记为0，为之后的生产过程中提供更加完整的产品数据信息。During the operation of the device, as shown in Figure 2, the schematic diagram of the installation position of the hardware acquisition equipment module of the device, the imaging diagram of the camera acquisition module of the Figure 3 device, the flowchart of the camera module triggered by the software in Figure 4, and the schematic diagram of the exterior structure of the device's house number in Figure 7, according to the camera acquisition The position of the hardware between module 3 and infrared radiation module 4, when the product where the embossed house number 16 characters are located is transmitted through the production line, through the camera acquisition module 3, and enters the imaging area of camera 1 No. 10 and camera 2 No. 11, It is necessary to ensure that the embossed characters of the current house number 16 are in the center of the imaging area of camera No. 1 10 and camera 2 No. 11: firstly, the state of the infrared tube in the infrared interfacing module 4 is triggered to change, and secondly, through the software processing program module 2 The software triggers the camera module 6 to detect and judge whether there is a circle that meets the set size in the current acquisition picture. When it is detected that there is a unique, clear circle that meets the set size in the current acquisition picture, it indicates that the current If the acquisition area is valid, the centralized control module 5 will collect and take pictures of the current character image through soft triggering; at the same time, it will detect the on-off status of No. 12 of the infrared interfacing module and No. 13 of the infrared interfacing module: When No. 12 of the module 1 and No. 13 of the infrared cross-radiation module 2 are in the conduction state at the same time, it is judged that the product type where the 16 characters of the current house number is located is the door frame 14; in other states, it is judged that the product type where the 16 characters of the current house number is located is the facade 15, so as to determine the product type Carry out the distinction, and save and record the image currently collected, for the door frame 14 and the door face 15, after distinguishing the type, for the user's convenience and intuitive distinction, in the third place of the file name, the door frame 14 is recorded as 1 , the facade 15 is recorded as 0, to provide more complete product data information for the subsequent production process.
经硬件采集设备模块1采集得到的压印门牌16字符图像，首先经集控模块5调取得到字符图像，经软件处理程序模块2中所述的图像预处理模块7，完成对字符图像的处理加工以及字符位置的准确定位，得到不含背景像素干扰的纯字符图像。如图5装置图像预处理模块流程图所示，该模块主要处理步骤如下：The embossed house plate 16 character images collected by the hardware acquisition device module 1 are first transferred through the centralized control module 5 to obtain the character images, and the image preprocessing module 7 described in the software processing program module 2 is used to complete the processing of the character images Processing and accurate positioning of character positions to obtain pure character images without background pixel interference. As shown in the flowchart of the image preprocessing module of the device in Figure 5, the main processing steps of this module are as follows:
步骤一：对于采集得到的字符图像首先进行图像灰度化，在减少图像处理数据量的同时保证图像中的基本信息不丢失；其次进行高斯滤波与中值滤波，对图像中的噪声点进行初步去除，进一步提高图像质量；之后利用sobel边缘检测算子对图像X方向上进行梯度运算；之后利用阈值分割进行图像二值化，将字符图像转化为黑白二值图像，去除部分背景，获得字符区域的初步轮廓；最后对二值化之后的图像进行形态学运算，消除字符中的细小断裂区域以及细小的噪声点，将短小像素点进行连接合并，从而进一步去除非字符区域的干扰。Step 1: For the acquired character image, the image is grayscaled first, and the basic information in the image is not lost while reducing the amount of image processing data; secondly, Gaussian filtering and median filtering are performed to initially filter the noise points in the image Remove to further improve the image quality; then use the sobel edge detection operator to perform gradient calculations on the X direction of the image; then use threshold segmentation to perform image binarization, convert the character image into a black and white binary image, remove part of the background, and obtain the character area The preliminary outline; finally, the morphological operation is performed on the binarized image to eliminate the small broken areas and small noise points in the characters, and the short pixels are connected and merged, so as to further remove the interference of the non-character area.
步骤二：进行门牌16字符轮廓的检测与定位：首先通过边缘轮廓搜索函数函数，大致搜索字符图像的轮廓区域，返回轮廓边界的位置坐标矩阵contours，进而求取其中的最小包围矩阵；其次对获得的外围轮廓矩阵，利用漫水填充算法，可以较好的去除字符图像背景中的非字符区域，通过设置可连通像素的上下限以及连通方式来达到填充的效果，从而只保留门牌16字符区域。Step 2: Detection and positioning of the character outline of house number 16: firstly, through the edge contour search function, roughly search the contour area of the character image, return the position coordinate matrix contours of the contour boundary, and then obtain the minimum enclosing matrix; secondly, obtain By using the flood filling algorithm, the non-character area in the background of the character image can be better removed, and the filling effect can be achieved by setting the upper and lower limits of the connected pixels and the connection method, so that only the 16-character area of the house number is reserved.
步骤三：利用边缘检测Canny算子对字符图像做边缘检测，获得字符图像的大致边缘轮廓；其次，通过霍夫变换函数检测图中满足特定要求的直线段，检测区域为整幅图像的1/3至2/3区域，对于该区域中依次绘制出利用霍夫变换函数检测到的线段进行筛选，对于长度不满足基本像素点长度要求的线段舍弃。经过霍夫变换检测直线之后，图像中只保留下长度大于一定像素个数且两条直线之间间隔大于设定阈值(此处阈值可以根据经验值做调整，本算法中设置为20个像素)的直线段，并对这些最终检测得到的直线线段按照坐标位置进行排序，按照直线所在位置，顺序返回每条直线的起点与终点的X、Y坐标，根据起点终点坐标，最终可以确定字符所在区域的边界线，记做图像dst，最终可以获得门牌16字符的确定位置。Step 3: Use the edge detection Canny operator to perform edge detection on the character image to obtain the approximate edge contour of the character image; secondly, use the Hough transform function to detect the straight line segment that meets the specific requirements in the image, and the detection area is 1/2 of the entire image In the 3 to 2/3 area, the line segments detected by the Hough transform function are sequentially drawn in this area for screening, and the line segments whose length does not meet the basic pixel length requirements are discarded. After the straight line is detected by the Hough transform, only the length greater than a certain number of pixels and the interval between the two straight lines are greater than the set threshold are retained in the image (the threshold here can be adjusted according to the empirical value, which is set to 20 pixels in this algorithm) The straight line segments of these final detections are sorted according to the coordinate positions, and the X and Y coordinates of the starting point and the end point of each straight line are returned in order according to the position of the straight line. According to the coordinates of the starting point and the end point, the area where the character is located can be finally determined The boundary line of , recorded as the image dst, can finally obtain the definite position of the 16 characters of the house number.
步骤四：对字符图像dst，首先计算字符图像上下边界两直线的斜率，分别记为K1，K2，对于K1，K2，求其均值，记为K；其次，以字符图像区域的中心点坐标为旋转中心，将字符图像dst进行旋转，调用旋转函数，其中参数degree，即旋转角度通过K求其三角变换，计算其对应的角度得到，得到旋转之后的图像，记为图像dst2，目的在于将字符图像旋转成水平方向，保证字符为水平方向，方便后续的字符切分，再进行图像保存。Step 4: For the character image dst, first calculate the slope of the two straight lines on the upper and lower boundaries of the character image, which are recorded as K1 and K2 respectively. For K1 and K2, find their mean value and record it as K; secondly, take the coordinates of the center point of the character image area as The rotation center rotates the character image dst, and calls the rotation function. The parameter degree, that is, the rotation angle is obtained by calculating its triangular transformation through K and calculating the corresponding angle. The rotated image is recorded as image dst2, and the purpose is to convert the character The image is rotated to the horizontal direction to ensure that the characters are in the horizontal direction, which is convenient for subsequent character segmentation, and then the image is saved.
经以上处理步骤之后，可获得完整且没有背景干扰因素的纯字符图像区域，且字符压印顺序符合常见的文字顺序，方向为水平方向排列，为之后的字符切分模块8提供了图像数据。After the above processing steps, a complete pure character image area without background interference factors can be obtained, and the character embossing order conforms to the common character order, and the direction is arranged in the horizontal direction, which provides image data for the subsequent character segmentation module 8 .
经软件处理程序模块2中所述的图像预处理模块7处理得到的字符图像，如图6装置字符切分模块流程图所示，对经过图像预处理模块7之后的字符图像，首先以图像的左上角作为像素搜索的起点，对于整张图片，从上至下逐行扫描判断图像中每一点的像素值：首先判断该点像素值是否为0，若该点像素值为0(表示该点为黑色)，则继续扫描下一点像素值；若该点像素值为1(表示该点为白色)，则判断该点纵坐标是否处于图片整体宽度的1/3到2/3之间，对于纵坐标不满足要求的则返回判断条件1，进行下一个点的判断；若满足要求，则进行判断3：取该点纵坐标为基准，向左向右各30个像素范围内的点，统计该范围内点的像素值为1的像素点个数，判断满足要求的像素点个数是否大于该区域所有像素点个数的80％，若不满足要求，则返回判断条件1，进行下一个点的判断；若满足要求则返回该点的坐标值，记为点P1(X1，Y1)。同样的判断流程对于整张图片分别从上至下，从左至右扫描，得到精确的字符所在位置的四个点的坐标值，记为P2(X2，Y2)；P3(X3，Y3)；P4(X4，Y4)；其次，根据四个点的坐标值，求出字符的具体位置，框定矩形框，进一步缩小字符位置所在范围；之后，对确定精确外边框之后的字符图像，按照先验知识，依次分割单个字符。首先，分割汉字字符：按照W3，H3(即压印模型中汉字字符的宽度与长度)的长宽数据，在字符图像上以W3为标准，先确定第一个字符，即汉字字符的大致右边界；再以汉字字符的大致右边界为起点，从右向左逐列扫描字符图像，遍历每一列字符图像中的像素值，当且仅当该点像素值为1(表示该点为白色)且取该点横坐标为基准，向上向下各30个像素范围内的点，统计该范围内点的像素值为1的像素点个数，满足要求的像素点个数大于该区域所有像素点个数的80％时，记录当前点的横坐标，以该横坐标为轴，所在直线即为汉字字符的精确右边界。由此即可得到第一个字符，即汉字字符的精确位置；其次，对于之后的字母以及数字随机构成的6个字符，同样根据W1，H1、W2，H2的大小宽度，先确定字符的大致右边界，再从大致右边界自右向左逐列扫描字符图像，遍历每一列字符图像中的像素值，进而确定字符的精确右边界，由此依次得到每一个字符的精确位置；对于精确分割之后的单个字符，最后将其大小归一化为28*28像素的图片，此处进行大小归一化的目的在于后续所采用的网络模型要求其输入数据大小为28*28，，将经过大小归一化之后的单个字符记为切分结果图，切分结果图按照对应的单个字符的内容保存在对应的文件夹下，字符内容与文件夹名一一对应。The character image obtained through the image preprocessing module 7 described in the software processing program module 2 is processed, as shown in the flow chart of the device character segmentation module in Figure 6, to the character image after the image preprocessing module 7, first use the image The upper left corner is used as the starting point for pixel search. For the entire picture, scan from top to bottom to determine the pixel value of each point in the image: first, determine whether the pixel value of the point is 0, if the pixel value of the point is 0 (indicating that the point is black), then continue to scan the pixel value of the next point; if the pixel value of the point is 1 (indicating that the point is white), then judge whether the vertical coordinate of the point is between 1/3 and 2/3 of the overall width of the picture, for If the ordinate does not meet the requirements, return to Judgment Condition 1 to judge the next point; if it meets the requirements, go to Judgment 3: take the ordinate of the point as the reference, and count the points within 30 pixels from left to right. The number of pixels with a pixel value of 1 in the range, judge whether the number of pixels that meet the requirements is greater than 80% of the number of all pixels in the area, if not, return to judgment condition 1, and proceed to the next step Judgment of the point; if the requirements are met, the coordinate value of the point is returned, which is recorded as point P1 (X1, Y1). The same judgment process scans the entire picture from top to bottom and from left to right to obtain the coordinate values of the four points where the precise character is located, which is recorded as P2 (X2, Y2); P3 (X3, Y3); P4(X4, Y4); Secondly, according to the coordinate values of the four points, the specific position of the character is obtained, and the rectangular frame is framed to further narrow the range of the character position; after that, for the character image after the precise outer frame is determined, according to the prior knowledge, sequentially segmenting individual characters. First, segment the Chinese characters: according to the length and width data of W3 and H3 (that is, the width and length of the Chinese characters in the embossing model), use W3 as the standard on the character image to first determine the first character, that is, the approximate right of the Chinese character Boundary; then start from the approximate right boundary of the Chinese character, scan the character image column by column from right to left, and traverse the pixel values in each column of the character image, if and only if the pixel value of the point is 1 (indicating that the point is white) And take the abscissa of the point as the benchmark, count the points within the range of 30 pixels up and down, and count the number of pixels with a pixel value of 1 in the range, and the number of pixels that meet the requirements is greater than all the pixels in the area When the number is 80%, record the abscissa of the current point, take the abscissa as the axis, and the straight line where it is located is the exact right boundary of the Chinese character. From this, the first character, that is, the precise position of the Chinese character can be obtained; secondly, for the subsequent letters and numbers randomly composed of 6 characters, also according to the size and width of W1, H1, W2, H2, first determine the approximate character The right boundary, and then scan the character image column by column from right to left from the approximate right boundary, traverse the pixel values in each column of the character image, and then determine the precise right boundary of the character, thereby obtaining the precise position of each character in turn; for precise segmentation After the single character, its size is finally normalized into a picture of 28*28 pixels. The purpose of size normalization here is that the subsequent network model requires its input data size to be 28*28. After the size A single character after normalization is recorded as a segmentation result graph, and the segmentation result graph is stored in the corresponding folder according to the content of the corresponding single character, and the character content corresponds to the folder name one by one.
根据以上所述步骤，经过字符切分模块8之后最终即可得切分后的单个字符图像结果，并且按照字符内容一一对应保存在相对应的文件夹下，形成相应的切分结果图，为之后的字符识别提供数据。According to the steps described above, after the character segmentation module 8, the single character image result after segmentation can be finally obtained, and stored in the corresponding folder according to the character content one by one, forming a corresponding segmentation result map, Provide data for subsequent character recognition.
经软件处理程序模块2中所述的字符切分模块8处理得到的单个字符图像结果，经caffe学习与识别模块借助caffe与LeNet-5手写字体识别网络，处理完成网络模型的学习，并测试当前网络模型的好坏，最后将待识别字符与训练学习得到的网络相结合，完成最后的字符内容的识别工作，具体流程如下：The single character image result processed by the character segmentation module 8 described in the software processing program module 2 is processed by the caffe learning and recognition module with the help of caffe and LeNet-5 handwritten font recognition network to complete the learning of the network model, and test the current Whether the network model is good or bad, finally combine the characters to be recognized with the network obtained from training and learning to complete the final recognition of character content. The specific process is as follows:
首先，在该装置进行字符识别之前，需要对网络模型进行学习，需大量的样本输入，因此对于训练切分结果图先进行人工分类，将对应字符分别保存到对应命名的文件夹下，形成训练样本集，对于测试切分结果图进行人工分类，将对应字符分别保存到对应命名的文件夹下，形成测试样本集。First of all, before the device performs character recognition, the network model needs to be learned, and a large number of sample inputs are required. Therefore, manual classification is performed on the training segmentation result graph, and the corresponding characters are saved in correspondingly named folders to form a training model. For the sample set, manually classify the test segmentation result graph, and save the corresponding characters in the corresponding named folders to form a test sample set.
其次，将训练样本集中的图片按照顺序打上标签，打乱顺序，形成标签文件，并记录标签文件所在路径，对于测试样本集中的图片做相同的操作，生成相应的标签文件。Secondly, label the pictures in the training sample set in order, shuffle the order, form a label file, and record the path of the label file, do the same operation for the pictures in the test sample set, and generate the corresponding label file.
之后，进行图片格式的转换，由于LeNet-5网络中要求输入的数据格式是ldb或者lmdb格式，因此，在图片数据处理的最后一步，需要进行图片格式的转换。Afterwards, convert the image format. Since the input data format required in the LeNet-5 network is ldb or lmdb format, the image format conversion is required in the last step of image data processing.
修改完毕之后，运行bat文件，处理完毕之后将会在设置的需要保存的对应路径下，生成相应的测试集与训练集文件夹，其中所包含的就是转换完成之后的训练集与测试集的数据。之后，运行bat文件，生成模型文件。After the modification is completed, run the bat file. After the processing is completed, the corresponding test set and training set folders will be generated under the corresponding path that needs to be saved, which contains the data of the training set and test set after conversion. . After that, run the bat file to generate the model file.
测试网络模型的优劣。为了测试所得网络模型对于其他非训练样本集中的图片识别效果如何，需利用测试样本集中的图片对网络进行测试，需要注意的是，采集得到的6万张训练样本与1万张测试样本是互相独立不重合的，因此，可以测试出该网络对不同数据的优劣性能。Test the strengths and weaknesses of network models. In order to test how well the obtained network model recognizes pictures in other non-training sample sets, it is necessary to use the pictures in the test sample set to test the network. It should be noted that the collected 60,000 training samples and 10,000 test samples are mutually exclusive Independent and non-overlapping, therefore, it is possible to test the pros and cons of the network for different data.
最后，在装置运行过程中，基于硬件采集设备模块1所采集到的待识别的字符图像，同样经过上述流程得到待识别切分结果图；将待识别切分结果图按照caffe学习与识别模块9中的学习模块进行处理，得到格式转换之后的数据集，记为待识别字符数据集；调用caffe学习与识别模块9中学习模块所得的模型文件lenet_iter_10000.caffemodel，将待识别字符数据放入模型中进行识别，最终输出与模型库中对比相似度概率值最高的对应字符，即当前字符的识别结果，进行组合输出，即为最终的字符识别结果，由此完成最终的字符识别。Finally, during the operation of the device, based on the character image to be recognized collected by the hardware acquisition device module 1, the segmentation result graph to be recognized is also obtained through the above process; the segmentation result graph to be recognized is obtained according to the caffe learning and recognition module 9 The learning module in is processed to obtain the data set after format conversion, which is recorded as the character data set to be recognized; call the model file lenet_iter_10000.caffemodel obtained from the learning module in caffe learning and recognition module 9, and put the character data to be recognized into the model Recognition is carried out, and the corresponding character with the highest similarity probability value compared with the model library is finally output, that is, the recognition result of the current character, and combined output is the final character recognition result, thereby completing the final character recognition.
以上所述仅为本发明专利的较佳实施例。The above descriptions are only preferred embodiments of the patent of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810044470.6ACN108229483A (en) | 2018-01-11 | 2018-01-11 | Based on the doorplate pressed characters identification device under caffe and soft triggering |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810044470.6ACN108229483A (en) | 2018-01-11 | 2018-01-11 | Based on the doorplate pressed characters identification device under caffe and soft triggering |
| Publication Number | Publication Date |
|---|---|
| CN108229483Atrue CN108229483A (en) | 2018-06-29 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810044470.6APendingCN108229483A (en) | 2018-01-11 | 2018-01-11 | Based on the doorplate pressed characters identification device under caffe and soft triggering |
| Country | Link |
|---|---|
| CN (1) | CN108229483A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740601A (en)* | 2019-01-04 | 2019-05-10 | 北京奇艺世纪科技有限公司 | A kind of embossing types recognition methods and device |
| CN110598681A (en)* | 2019-10-22 | 2019-12-20 | 广东技术师范大学 | Natural street view house number identification method |
| CN110852419A (en)* | 2019-11-08 | 2020-02-28 | 中山大学 | An action model based on deep learning and its training method |
| CN112861736A (en)* | 2021-02-10 | 2021-05-28 | 上海大学 | Document table content identification and information extraction method based on image processing |
| CN113378819A (en)* | 2021-06-24 | 2021-09-10 | 成都飞机工业（集团）有限责任公司 | Rapid identification method for airborne software version |
| CN114022453A (en)* | 2021-11-07 | 2022-02-08 | 苏州众智图优智能科技有限公司 | Screen printing plate pattern detection method and detection equipment for solar cell |
| CN114067192A (en)* | 2022-01-07 | 2022-02-18 | 北京许先网科技发展有限公司 | Character recognition method and system |
| CN114140796A (en)* | 2021-11-30 | 2022-03-04 | 马鞍山学院 | Shaft part surface character visual identification method based on linear array camera |
| CN116612476A (en)* | 2023-07-20 | 2023-08-18 | 合肥图迅电子科技有限公司 | Chip character detection method and device, computer equipment and storage medium |
| CN117710985A (en)* | 2023-12-18 | 2024-03-15 | 珠海凌烟阁芯片科技有限公司 | Optical character recognition method and device and intelligent terminal |
| CN120046118A (en)* | 2025-04-23 | 2025-05-27 | 江苏时代新能源科技有限公司 | Tab welding detection method, device, storage medium and program product |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1684096A (en)* | 2004-04-16 | 2005-10-19 | 中国科学院自动化研究所 | House number automatic recognition system and method |
| CN101303803A (en)* | 2008-06-11 | 2008-11-12 | 北京中星微电子有限公司 | Method and system for discriminating license plate |
| US20090028435A1 (en)* | 2007-07-23 | 2009-01-29 | Bo Wu | Character image extracting apparatus and character image extracting method |
| CN101697196A (en)* | 2009-10-29 | 2010-04-21 | 上海索广电子有限公司 | Digital identification system and method for serial numbers of name plate of camera |
| CN104408931A (en)* | 2014-10-29 | 2015-03-11 | 合肥指南针电子科技有限责任公司 | Incomplete sign license plate identification system and method |
| CN104871180A (en)* | 2012-12-13 | 2015-08-26 | 高通股份有限公司 | Text image quality based feedback for OCR |
| CN105975968A (en)* | 2016-05-06 | 2016-09-28 | 西安理工大学 | Caffe architecture based deep learning license plate character recognition method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1684096A (en)* | 2004-04-16 | 2005-10-19 | 中国科学院自动化研究所 | House number automatic recognition system and method |
| US20090028435A1 (en)* | 2007-07-23 | 2009-01-29 | Bo Wu | Character image extracting apparatus and character image extracting method |
| CN101303803A (en)* | 2008-06-11 | 2008-11-12 | 北京中星微电子有限公司 | Method and system for discriminating license plate |
| CN101697196A (en)* | 2009-10-29 | 2010-04-21 | 上海索广电子有限公司 | Digital identification system and method for serial numbers of name plate of camera |
| CN104871180A (en)* | 2012-12-13 | 2015-08-26 | 高通股份有限公司 | Text image quality based feedback for OCR |
| CN104408931A (en)* | 2014-10-29 | 2015-03-11 | 合肥指南针电子科技有限责任公司 | Incomplete sign license plate identification system and method |
| CN105975968A (en)* | 2016-05-06 | 2016-09-28 | 西安理工大学 | Caffe architecture based deep learning license plate character recognition method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740601A (en)* | 2019-01-04 | 2019-05-10 | 北京奇艺世纪科技有限公司 | A kind of embossing types recognition methods and device |
| CN110598681A (en)* | 2019-10-22 | 2019-12-20 | 广东技术师范大学 | Natural street view house number identification method |
| CN110852419A (en)* | 2019-11-08 | 2020-02-28 | 中山大学 | An action model based on deep learning and its training method |
| CN110852419B (en)* | 2019-11-08 | 2023-05-23 | 中山大学 | An action model based on deep learning and its training method |
| CN112861736B (en)* | 2021-02-10 | 2022-08-09 | 上海大学 | Document table content identification and information extraction method based on image processing |
| CN112861736A (en)* | 2021-02-10 | 2021-05-28 | 上海大学 | Document table content identification and information extraction method based on image processing |
| CN113378819A (en)* | 2021-06-24 | 2021-09-10 | 成都飞机工业（集团）有限责任公司 | Rapid identification method for airborne software version |
| CN113378819B (en)* | 2021-06-24 | 2022-11-18 | 成都飞机工业（集团）有限责任公司 | Rapid identification method for airborne software version |
| CN114022453A (en)* | 2021-11-07 | 2022-02-08 | 苏州众智图优智能科技有限公司 | Screen printing plate pattern detection method and detection equipment for solar cell |
| CN114140796A (en)* | 2021-11-30 | 2022-03-04 | 马鞍山学院 | Shaft part surface character visual identification method based on linear array camera |
| CN114140796B (en)* | 2021-11-30 | 2024-09-27 | 马鞍山学院 | Shaft part surface character visual recognition method based on linear array camera |
| CN114067192A (en)* | 2022-01-07 | 2022-02-18 | 北京许先网科技发展有限公司 | Character recognition method and system |
| CN116612476A (en)* | 2023-07-20 | 2023-08-18 | 合肥图迅电子科技有限公司 | Chip character detection method and device, computer equipment and storage medium |
| CN116612476B (en)* | 2023-07-20 | 2023-09-22 | 合肥图迅电子科技有限公司 | Chip character detection method and device, computer equipment and storage medium |
| CN117710985A (en)* | 2023-12-18 | 2024-03-15 | 珠海凌烟阁芯片科技有限公司 | Optical character recognition method and device and intelligent terminal |
| CN120046118A (en)* | 2025-04-23 | 2025-05-27 | 江苏时代新能源科技有限公司 | Tab welding detection method, device, storage medium and program product |
| Publication | Publication Date | Title |
|---|---|---|
| CN108229483A (en) | Based on the doorplate pressed characters identification device under caffe and soft triggering | |
| CN101916364B (en) | Adaptive dial digital identification method | |
| CN107545239B (en) | Fake plate detection method based on license plate recognition and vehicle characteristic matching | |
| CN107257161B (en) | Transformer substation disconnecting link remote control auxiliary checking method and system based on state recognition algorithm | |
| CN114549981A (en) | A deep learning-based intelligent inspection pointer meter identification and reading method | |
| CN102704215B (en) | Automatic cutting method of embroidery cloth based on combination of DST file parsing and machine vision | |
| WO2018018788A1 (en) | Image recognition-based meter reading apparatus and method thereof | |
| CN105046252A (en) | Method for recognizing Renminbi (Chinese currency yuan) crown codes | |
| CN111695373B (en) | Zebra stripes positioning method, system, medium and equipment | |
| CN104392205A (en) | Abnormal vehicle license plate recognition method and system | |
| CN107895492A (en) | A kind of express highway intelligent analysis method based on conventional video | |
| Soomro et al. | Vehicle number recognition system for automatic toll tax collection | |
| CN108615058A (en) | A kind of method, apparatus of character recognition, equipment and readable storage medium storing program for executing | |
| CN106991421A (en) | A kind of ID card information extraction system | |
| CN110032932B (en) | Human body posture identification method based on video processing and decision tree set threshold | |
| CN110599463B (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network | |
| CN103440035A (en) | Gesture recognition system in three-dimensional space and recognition method thereof | |
| CN101584624A (en) | Guideboard recognition blind-guide device and method thereof based on DSP | |
| CN103699876B (en) | Method and device for identifying vehicle number based on linear array CCD (Charge Coupled Device) images | |
| CN114596242A (en) | Defect detection method, apparatus, electronic device, and computer-readable storage medium | |
| CN113076860B (en) | Bird detection system under field scene | |
| CN112784835A (en) | Method and device for identifying authenticity of circular seal, electronic equipment and storage medium | |
| CN108362708A (en) | Pen core scribing line standardized detection system based on machine vision and method | |
| CN116402822B (en) | Concrete structure image detection method, device, electronic equipment and storage medium | |
| CN103886319A (en) | Intelligent held board recognizing method based on machine vision |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | Application publication date:20180629 |