CN110390341B - Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods - Google Patents
Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goodsDownload PDFInfo
- Publication number
- CN110390341B CN110390341BCN201910683531.8ACN201910683531ACN110390341BCN 110390341 BCN110390341 BCN 110390341BCN 201910683531 ACN201910683531 ACN 201910683531ACN 110390341 BCN110390341 BCN 110390341B
- Authority
- CN
- China
- Prior art keywords
- image
- data
- green traffic
- convolutional neural
- cargo
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
Description
Translated fromChinese技术领域technical field
本发明属于图像识别技术领域，特别涉及基于卷积神经网络的绿通车运载货物放射源图像识别方法。The invention belongs to the technical field of image recognition, in particular to a convolutional neural network-based image recognition method for radioactive sources of goods carried by green traffic.
背景技术Background technique
放射源检查通常采用γ射线和x射线，当射线穿透该物体时，探测器最终检测的能量是该方向上射线被多种物质吸收后的结果，此时的衰减系数是一个关于空间位置和能量的函数。为获得不同空间位置的能量，需要对被检物体进行多方位投影，得到各方向的最终能量值。经反Radon变换后即可得到各空间位置的能量值。将能量值映射为像素值，经重建计算获得图像矩阵。放射源成像质量影响因素包括两个方面，一是硬件条件，主要有设备焦点尺寸、成像放大倍数、管电压、电流等[33]。二是被检测物自身属性，主要有被检测物结构复杂度、被检测物自身尺寸、被检测物的构成物质密度及厚度差异等。Radiation source inspection usually uses gamma rays and x-rays. When the rays penetrate the object, the energy finally detected by the detector is the result of the rays being absorbed by various substances in this direction. The attenuation coefficient at this time is a function related to the spatial position and function of energy. In order to obtain the energy of different spatial positions, it is necessary to perform multi-directional projection on the inspected object to obtain the final energy value in each direction. After inverse Radon transformation, the energy value of each spatial position can be obtained. The energy value is mapped to the pixel value, and the image matrix is obtained through reconstruction calculation. Factors affecting the imaging quality of radioactive sources include two aspects. One is hardware conditions, which mainly include equipment focus size, imaging magnification, tube voltage, current, etc. [33]. The second is the properties of the detected object itself, mainly including the structural complexity of the detected object, the size of the detected object itself, the density and thickness difference of the constituent materials of the detected object, etc.
放射源检查在国内外众多行业得到广泛运用，包括安检、医疗、工业检测等领域，对放射源图像的自动识别成为亟需解决的热点问题，为提高放射源图像识别准确率，国内外的众多学者为此进行了大量的研究。绿通车运载货物放射源图像与医学、安检等领域的放射源图像特征较为相似。Radioactive source inspection is widely used in many industries at home and abroad, including security inspection, medical treatment, industrial inspection and other fields. The automatic recognition of radioactive source images has become a hot issue that needs to be solved urgently. In order to improve the accuracy of radioactive source image recognition, many domestic and foreign Scholars have done a lot of research on this. The characteristics of radioactive source images of goods carried by green traffic are similar to those of radioactive source images in medicine, security inspection and other fields.
基于模式识别与传统机器学习的图像分类算法并不适用于绿通车运载货物放射源图像分类问题。主要原因是针对放射源图像数据集，人工设定规则提取图像特征难度较大，并且模式识别与传统机器学习无法满足大数据集的计算复杂度要求。而卷积神经网络强大的图像特征提取能力及高效的计算能力使得基于卷积神经网络的图像分类模型在放射源图像分类方面有着惊人的表现，因此选用卷积神经网络作为基本算法，并对经典卷积神经网络常用模型进行改进。The image classification algorithm based on pattern recognition and traditional machine learning is not suitable for the problem of image classification of radioactive sources carried by green traffic. The main reason is that for radioactive source image datasets, it is difficult to manually set rules to extract image features, and pattern recognition and traditional machine learning cannot meet the computational complexity requirements of large datasets. The powerful image feature extraction ability and efficient computing ability of the convolutional neural network make the image classification model based on the convolutional neural network have amazing performance in the classification of radioactive source images. Therefore, the convolutional neural network is selected as the basic algorithm, and the classic Convolutional neural network commonly used models are improved.
卷积神经网络结构从数据流角度可分为输入层、隐藏层和输出层。在图像识别领域的应用中，输入层可直接处理图像原始像素值，并对输入数据进行标准化处理，以提高模型性能。隐藏层的基本构造中包含卷积层、池化层和全连接层，该层可对输入数据进行降维处理，实现图像特征自动提取，并通过全连接层实现分类任务。输出层负责对上层的分类任务进行结果输出。卷积神经网络可在不改变图像网络拓扑结构的基础上，可实现图像信息特征自动提取。相对于其他神经网络结构，卷积神经网络具有局部连接和权值共享的特性，能够减少网络训练时所需参数个数，具有强大的高维数据处理能力。The convolutional neural network structure can be divided into input layer, hidden layer and output layer from the perspective of data flow. In applications in the field of image recognition, the input layer can directly process the original pixel values of the image and standardize the input data to improve model performance. The basic structure of the hidden layer includes a convolutional layer, a pooling layer, and a fully connected layer. This layer can perform dimensionality reduction processing on the input data, realize automatic extraction of image features, and implement classification tasks through the fully connected layer. The output layer is responsible for outputting the results of the classification tasks of the upper layer. The convolutional neural network can realize the automatic extraction of image information features without changing the topological structure of the image network. Compared with other neural network structures, the convolutional neural network has the characteristics of local connection and weight sharing, which can reduce the number of parameters required for network training and has powerful high-dimensional data processing capabilities.
基于卷积神经网络的放射源图像识别技术在医疗、安检、工业检测等领域得到广泛应用。但在高速公路绿通车检查领域，国内外学者未进行相应研究。The radioactive source image recognition technology based on convolutional neural network has been widely used in medical treatment, security inspection, industrial inspection and other fields. But in the field of expressway green traffic inspection, domestic and foreign scholars have not conducted corresponding research.
当前绿通车检验方式主要为人工开箱检查和放射源图像人工评判两种方式。人工开箱检查方式耗时费力，检查时间不可控，易引起交通拥堵。放射源图像人工评判是查验人员通过放射源扫描图像判断货物外形、灰度值和装载均匀性等是否符合货物特点，来判断车厢内部是否夹带有非法物品。人工评判耗时长、难度大且准确度较低，因此难以成为理想的检查方式。The current green traffic inspection methods are mainly two methods: manual unpacking inspection and manual evaluation of radioactive source images. The manual unpacking inspection method is time-consuming and laborious, the inspection time is uncontrollable, and it is easy to cause traffic congestion. The manual evaluation of radioactive source images is to judge whether the shape, gray value and loading uniformity of the goods conform to the characteristics of the goods through the scanning images of the radioactive sources, so as to judge whether there are illegal items inside the compartment. Manual evaluation is time-consuming, difficult, and less accurate, making it difficult to be an ideal inspection method.
发明内容Contents of the invention
本发明的目的在于提供基于卷积神经网络的绿通车运载货物放射源图像识别方法，以解决上述问题。The object of the present invention is to provide an image recognition method for radioactive sources of goods carried by green traffic based on convolutional neural network, so as to solve the above problems.
为实现上述目的，本发明采用以下技术方案：To achieve the above object, the present invention adopts the following technical solutions:
基于卷积神经网络的绿通车运载货物放射源图像识别方法，包括以下步骤：A method for recognizing images of radioactive sources of goods carried by green traffic based on convolutional neural networks, including the following steps:
步骤1，绿通车运载货物放射源图像预处理：对绿通车运载货物放射源图像进行数据清洗、货物区域分割、降噪处理、数据增强和通道转化；Step 1, preprocessing of the radioactive source image of the goods carried by the green traffic: data cleaning, cargo area segmentation, noise reduction processing, data enhancement, and channel conversion are performed on the radioactive source image of the cargo transported by the green traffic;
步骤2，输入图像样本准备：将图像尺寸大小调整为统一尺寸，并以8:1:1的数量比例分为训练集、验证集及测试集；Step 2, input image sample preparation: adjust the size of the image to a uniform size, and divide it into a training set, a verification set and a test set with a ratio of 8:1:1;
步骤3，设计绿通车运载货物放射源图像识别模型；Step 3, design the image recognition model of the radioactive source of goods carried by green traffic;
步骤4，对绿通车运载货物放射源图像识别模型进行调优：首先提高数据集图像质量，然后使用数据迭代器解决内存占用过大问题，增加网络输入图像尺寸；Step 4. Tuning the image recognition model of radioactive sources carried by green traffic vehicles: first, improve the image quality of the data set, and then use the data iterator to solve the problem of excessive memory usage and increase the size of the network input image;
步骤5，通过对模型的训练、验证及测试，并记录训练过程中损失函数及分类准确率的变化，损失函数能够反映模型对货物类型准确分类的能力。Step 5, through the training, verification and testing of the model, and recording the changes in the loss function and classification accuracy during the training process, the loss function can reflect the ability of the model to accurately classify the types of goods.
进一步的，步骤1中，绿通车运载货物放射源图像数据清洗具体包括：Further, in step 1, the cleansing of the image data of the radioactive source of the goods carried by the green traffic specifically includes:
选用人工检查结果与放射源检查结果相一致的图像数据，通过匹配数据源中SCANGOODS和MANUALGOODS两个字段来完成该阶段数据清洗任务；Select the image data whose manual inspection results are consistent with the radioactive source inspection results, and complete the data cleaning task at this stage by matching the two fields SCANGOODS and MANUALGOODS in the data source;
删除数据源中货物区域占比过小的图像。Delete images with too small a proportion of the cargo area in the data source.
进一步的，步骤1中，绿通车运载货物放射源图像货物区域分割：Further, in step 1, the cargo region segmentation of the radioactive source image of the cargo carried by the green traffic:
确定图像中货物区域的行、列下标公式如下：The formula for determining the row and column subscripts of the cargo area in the image is as follows:
式中rowmin为货物区域的行下标最小值，rowmax为货物区域行下标最大值，colmin为货物区域列下标最小值，colmax为货物区域列下标最大值，Xij为图像数据矩阵，row为数据矩阵的行数，col为数据矩阵的列数，where()函数表示满足条件的像素点下标值数组；In the formula, rowmin is the minimum value of the row subscript of the cargo area, rowmax is the maximum value of the row subscript of the cargo area, colmin is the minimum value of the column subscript of the cargo area, colmax is the maximum value of the column subscript of the cargo area, Xij is Image data matrix, row is the number of rows in the data matrix, col is the number of columns in the data matrix, and the where() function represents an array of pixel subscript values that meet the conditions;
利用上述公式对原始图像数据进行货物区域分割。Cargo area segmentation is performed on the original image data using the above formula.
进一步的，步骤1中，绿通车运载货物放射源图像降噪处理：采用中值滤波器，将图像中的各像素值替换为窗口区域内像素值的中值，降低图像中的噪声污染。Further, in step 1, the noise reduction processing of the image of the radioactive source carrying goods by green traffic: using a median filter to replace each pixel value in the image with the median value of the pixel values in the window area to reduce the noise pollution in the image.
进一步的，步骤1中，利用翻转、旋转等方法对绿通车运载货物放射源图像进行数据增强处理。Further, in step 1, data enhancement processing is performed on the image of the radioactive source of the goods carried by the green traffic by using methods such as flipping and rotating.
进一步的，步骤1中，将单通道图像过渡为三通道图像，图像表征无变化，达到符合网络输入数据格式要求的目的，采用OpenCV中COLOR_GRAY2BGR函数实现由单通道数据转换为三通道数据的功能，转换公式如下：Further, in step 1, the single-channel image is transformed into a three-channel image, and the image representation does not change, so as to meet the requirements of the network input data format. The COLOR_GRAY2BGR function in OpenCV is used to realize the function of converting single-channel data into three-channel data. The conversion formula is as follows:
B＝G＝R＝GrayB=G=R=Gray
即单通道图的像素点分布在RGB空间的直线R＝G＝B上，使用该函数对绿通车运载货物放射源图像进行图像通道转换。That is, the pixels of the single-channel image are distributed on the straight line R=G=B in the RGB space, and this function is used to perform image channel conversion on the image of the radioactive source of the green traffic.
进一步的，步骤3中，基于随机森林思想将多种常用卷积神经网络模型输出结果的平均值作为网络的最终输出结果，通过集成互不依赖的多种网络模型，以并行运算的方式，通过“投票法”得出输出结果；采用AlexNet、GoogLeNet、ResNet和DenseNet四种卷积神经网络常用模型建立集成学习规则，对上述四种卷积神经网络常用模型预测结果进行混合运算，取平均预测概率最大值对应货物类型为最终结果；Further, in step 3, based on the idea of random forest, the average value of the output results of various commonly used convolutional neural network models is used as the final output result of the network. "Voting method" to obtain the output results; four common models of convolutional neural networks, AlexNet, GoogLeNet, ResNet and DenseNet, are used to establish integrated learning rules, and the prediction results of the above four common models of convolutional neural networks are mixed and calculated, and the average prediction probability is taken The cargo type corresponding to the maximum value is the final result;
以Cifar10数据集对AlexNet、DenseNet、GoogleNet和ResNet-进行预训练，然后使用绿通车运载货物放射源图像进行重新训练以微调模型。AlexNet, DenseNet, GoogleNet, and ResNet- were pre-trained on the Cifar10 dataset, and then re-trained using the radioactive source images of the green traffic to fine-tune the model.
进一步的，步骤4中，采用数据迭代器对程序进行优化，迭代器按需读取数据，解决内存爆炸问题，将输入图像尺寸直接增加至AlexNet、GoogLeNet、ResNet和DenseNet模型规定尺寸大小，并分别设置实验。Further, in step 4, the program is optimized using the data iterator, which reads data on demand, solves the problem of memory explosion, directly increases the size of the input image to the specified size of the AlexNet, GoogLeNet, ResNet and DenseNet models, and respectively Set up the experiment.
与现有技术相比，本发明有以下技术效果：Compared with the prior art, the present invention has the following technical effects:
本发明采用基于卷积神经网络的绿通车运载货物放射源图像识别可避免查验结果对查验人员的主观依赖性，减轻一线查验人员工作强度。同时可提高检查效率，减少收费站拥堵情况的产生；The present invention adopts the image recognition of the radioactive source of the goods carried by green traffic based on the convolutional neural network, which can avoid the subjective dependence of the inspection results on the inspection personnel, and reduce the work intensity of the front-line inspection personnel. At the same time, it can improve the inspection efficiency and reduce the congestion of toll stations;
不同绿通车放射源扫描设备的放射源、穿透力等参数不同，会导致同一货物在不同品牌绿通车放射源检查设备下的扫面成像外观表现也不同，采用本发明可有效避免稽查人员的误判行为；The radioactive source and penetrating power of different green traffic radioactive source scanning equipment are different, which will lead to different appearances of the same cargo in scanning imaging under different brands of green traffic radioactive source inspection equipment. The use of the present invention can effectively avoid inspectors’ confusion. Misjudgment;
本发明的研究结果在实现判断货物是否符合鲜活农产品目录的基础上，实现了绿通车运载货物类型的精准识别，为以后挖掘农产品流转规律提供前导性研究，同时为放射源检查装置实现货物自动识别提供理论参考；The research result of the present invention realizes the accurate identification of the type of goods carried by green traffic vehicles on the basis of judging whether the goods conform to the catalog of fresh and live agricultural products. Identification provides theoretical reference;
本发明的所有成果均依赖于数据源支撑，图像质量的优劣对模型性能有着至关重要的作用。由于原始图像数据存在标签错误、有效信息占比不足、噪声污染及分布不均匀等缺陷，依次采用数据清洗、货物区域分割、图像降噪及数据增强等方式对放射源图像进行预处理，大大提升了模型性能及模型泛化能力；All the achievements of the present invention depend on the support of the data source, and the quality of the image has a crucial effect on the performance of the model. Due to defects such as label errors, insufficient proportion of effective information, noise pollution, and uneven distribution of the original image data, the radioactive source image is preprocessed by sequentially adopting data cleaning, cargo area segmentation, image noise reduction, and data enhancement, which greatly improves the quality of the image. Improved model performance and model generalization capabilities;
本发明采用数据迭代器对程序进行优化，降低了程序内存占用率，有效解决了内存爆炸问题。The invention adopts the data iterator to optimize the program, reduces the memory occupation rate of the program, and effectively solves the memory explosion problem.
附图说明Description of drawings
图1为绿通车放射源图像人工评判流程。Figure 1 is the manual evaluation process of the radioactive source image of Green Traffic.
图2为基于卷积神经网络的绿通车运载货物放射源图像识别流程。Figure 2 shows the image recognition process of radioactive sources of goods carried by green traffic based on convolutional neural network.
图3为基于卷积神经网络的绿通车运载货物放射源图像识别的设计流程。Figure 3 shows the design process of the image recognition of radioactive sources of goods carried by green traffic based on convolutional neural network.
图4为RF-CNN模型图示。Figure 4 is an illustration of the RF-CNN model.
具体实施方式detailed description
以下结合附图对本发明进一步说明：The present invention is further described below in conjunction with accompanying drawing:
请参阅图1至图4，基于卷积神经网络的绿通车运载货物放射源图像识别主要包括以下四个部分，分别为绿通车运载货物放射源图像预处理；绿通车运载货物放射源图像识别模型设计；绿通车运载货物放射源图像识别模型调优；验证实验结果。Please refer to Figures 1 to 4. The image recognition of radioactive sources of goods carried by green vehicles based on convolutional neural networks mainly includes the following four parts, which are image preprocessing of radioactive sources of goods carried by green vehicles; image recognition model of radioactive sources of goods carried by green vehicles Design; optimization of the radioactive source image recognition model for green traffic vehicles; verification of experimental results.
各部分具体如下：The details of each part are as follows:
1、绿通车运载货物放射源图像预处理1. Preprocessing of images of radioactive sources carried by green vehicles
绿通车运载货物放射源图像预处理主要包括数据清洗、货物区域分割、降噪处理、数据增强和通道转化。The image preprocessing of radioactive source images carried by green traffic vehicles mainly includes data cleaning, cargo area segmentation, noise reduction processing, data enhancement and channel conversion.
绿通车运载货物放射源图像数据清洗Image data cleaning of radioactive sources carried by green vehicles
为了保证输入网络的图像数据标签可靠以及图像信息足够丰富，需要对绿通车运载货物放射源图像有效性进行分析。根据绿通车运载货物放射源图像特点和分析需求，需对数据源进行数据清洗，该过程分两步进行。In order to ensure that the image data labels input into the network are reliable and the image information is rich enough, it is necessary to analyze the validity of the radioactive source images of the goods carried by green traffic. According to the image characteristics and analysis requirements of the radioactive source of the goods carried by the Green Express, the data source needs to be cleaned, and the process is divided into two steps.
为保证数据标签可靠，只选用人工检查结果与放射源检查结果相一致的图像数据。通过匹配数据源中SCANGOODS和MANUALGOODS两个字段来完成该阶段数据清洗任务。In order to ensure the reliability of the data label, only the image data whose manual inspection results are consistent with the radioactive source inspection results are selected. The data cleaning task at this stage is completed by matching the two fields SCANGOODS and MANUALGOODS in the data source.
为保证所选图像数据的有效性，需删除数据源中货物区域占比过小的图像。In order to ensure the validity of the selected image data, it is necessary to delete the images in the data source whose proportion of the cargo area is too small.
绿通车运载货物放射源图像货物区域分割Segmentation of cargo area in radioactive source image of green traffic
由于车辆尺寸、车辆与放射源检查设备间距等原因，合格绿通车运载货物放射源图像中仍然存在大面积非货物区域。但由于放射源图像中存在噪声污染，图像中非货物区域并非完全空白。而绿通车运载货物放射源图像中非货物区域对提取图像中货物特征信息无任何有利作用，若直接将该类图像作为神经网络输入数据，将导致网络计算负担大幅度增加，所以货物区域分割是十分有必要的。Due to the size of the vehicle, the distance between the vehicle and the radioactive source inspection equipment, etc., there are still large areas of non-cargo areas in the images of radioactive sources of goods carried by qualified green traffic vehicles. However, due to noise pollution in the image of radioactive sources, the non-cargo area in the image is not completely blank. However, the non-cargo area in the image of the radioactive source of the green traffic carrying goods has no beneficial effect on the extraction of the characteristic information of the goods in the image. If this type of image is directly used as the input data of the neural network, the network calculation burden will be greatly increased. Therefore, the segmentation of the cargo area is very important. It is very necessary.
经过多次实验，最终确定图像中货物区域的行、列下标公式如下：After many experiments, the row and column subscript formulas of the cargo area in the image are finally determined as follows:
式中rowmin为货物区域的行下标最小值，rowmax为货物区域行下标最大值，colmin为货物区域列下标最小值，colmax为货物区域列下标最大值，Xij为图像数据矩阵，row为数据矩阵的行数，col为数据矩阵的列数，where()函数表示满足条件的像素点下标值数组。In the formula, rowmin is the minimum value of the row subscript of the cargo area, rowmax is the maximum value of the row subscript of the cargo area, colmin is the minimum value of the column subscript of the cargo area, colmax is the maximum value of the column subscript of the cargo area, Xij is Image data matrix, row is the number of rows in the data matrix, col is the number of columns in the data matrix, and the where() function represents an array of pixel subscript values that meet the conditions.
利用卷积神经网络进行图像识别，对硬件设备的计算能力要求较高，利用上述公式对原始图像数据进行货物区域分割，可进一步增强图像信息。将有限的计算资源最大化的利用在有效信息的提取方面。同时，增加图像数据的有效信息占比，能够去除无关信息对识别结果的影响，提高模型识别准确率。The use of convolutional neural networks for image recognition requires high computing power for hardware devices. Using the above formula to segment the original image data into cargo areas can further enhance image information. Maximize the use of limited computing resources in the extraction of effective information. At the same time, increasing the effective information ratio of image data can remove the influence of irrelevant information on the recognition results and improve the accuracy of model recognition.
绿通车运载货物放射源图像降噪处理Noise reduction processing of radioactive source images carried by green traffic
放射源图像在生成、传输、接收和处理的过程中，不可避免的存在外部干扰或内部干扰，如探测器中敏感元件灵敏度的不均匀性，数字化过程中产生量化误差，传输过程中的误差以及人为因素等。In the process of generation, transmission, reception and processing of radioactive source images, external or internal interference inevitably exists, such as the inhomogeneity of the sensitivity of sensitive elements in the detector, quantization errors in the digitization process, errors in the transmission process, and human factors etc.
本发明采用中值滤波器，其原理是将图像中的各像素值替换为窗口区域内像素值的中值。与其他滤波方式相比，采用中值滤波降噪后的图像峰值信噪比最大，能最大限度的降低图像中的噪声污染。The present invention adopts the median filter, and its principle is to replace each pixel value in the image with the median value of the pixel values in the window area. Compared with other filtering methods, the peak signal-to-noise ratio of the image after noise reduction using the median filter is the largest, which can minimize the noise pollution in the image.
在绿通车运载货物放射源图像中，各类数据分布不均匀，为了最大限度地增加训练样本的数量，同时均衡样本在各个类别的分布，在图像分类问题中，数据增强经常被使用。数据增强的本质是利用有限的数据生成更多的等价数据，只改变训练数据的数组表征而保持标签不变。增加训练数据的数量，防止过度拟合，同时提高网络模型的泛化能力。In the images of radioactive sources carrying goods in green traffic, various types of data are distributed unevenly. In order to maximize the number of training samples and balance the distribution of samples in each category, data augmentation is often used in image classification problems. The essence of data augmentation is to use limited data to generate more equivalent data, and only change the array representation of the training data while keeping the labels unchanged. Increase the amount of training data, prevent overfitting, and improve the generalization ability of the network model.
利用翻转、旋转等方法对绿通车运载货物放射源图像进行数据增强处理。Using methods such as flipping and rotating, data enhancement processing is performed on the image of the radioactive source of the goods carried by the green traffic.
绿通车运载货物放射源图像为灰度图，即单通道图，每个像素点由0到255之间的单一数字表示灰度值。而论文采用的四种经典卷积神经网络模型的应用场景均为对自然图像的处理，自然图像为三通道图，每个像素点由三个数字表示颜色属性。为实现对经典卷积神经网络模型的迁移应用，需要对绿通车运载货物放射源图像进行通道转换处理。The image of the radioactive source of the goods carried by the green traffic is a grayscale image, that is, a single-channel image, and each pixel is represented by a single number between 0 and 255. Grayscale value. The application scenarios of the four classic convolutional neural network models used in this paper are all processing natural images. Natural images are three-channel images, and each pixel is represented by three numbers for color attributes. In order to realize the migration and application of the classic convolutional neural network model, it is necessary to perform channel conversion processing on the radioactive source images of the green traffic.
在图像通道转换研究中，将单通道图像简单过渡为三通道图像，图像表征无变化，达到符合网络输入数据格式要求的目的。采用OpenCV中COLOR_GRAY2BGR函数实现由单通道数据转换为三通道数据的功能，转换公式如下：In the research of image channel conversion, the single-channel image is simply transformed into a three-channel image, and the image representation does not change, so as to meet the requirements of the network input data format. Use the COLOR_GRAY2BGR function in OpenCV to realize the function of converting single-channel data to three-channel data. The conversion formula is as follows:
B＝G＝R＝GrayB=G=R=Gray
即单通道图的像素点分布在RGB空间的直线R＝G＝B上，使用该函数对绿通车运载货物放射源图像进行图像通道转换。That is, the pixels of the single-channel image are distributed on the straight line R=G=B in the RGB space, and this function is used to perform image channel conversion on the image of the radioactive source of the green traffic.
2、绿通车运载货物放射源图像识别模型设计2. Design of image recognition model for radioactive sources of goods carried by green traffic
卷积神经网络结构从数据流角度可分为输入层、隐藏层和输出层。在图像识别领域的应用中，输入层可直接处理图像原始像素值，并对输入数据进行标准化处理，以提高模型性能。隐藏层的基本构造中包含卷积层、池化层和全连接层，该层可对输入数据进行降维处理，实现图像特征自动提取，并通过全连接层实现分类任务。输出层负责对上层的分类任务进行结果输出。The convolutional neural network structure can be divided into input layer, hidden layer and output layer from the perspective of data flow. In applications in the field of image recognition, the input layer can directly process the original pixel values of the image and standardize the input data to improve model performance. The basic structure of the hidden layer includes a convolutional layer, a pooling layer, and a fully connected layer. This layer can perform dimensionality reduction processing on the input data, realize automatic extraction of image features, and implement classification tasks through the fully connected layer. The output layer is responsible for outputting the results of the classification tasks of the upper layer.
卷积神经网络在图像识别领域的应用不断深入，优秀算法层出不穷。通常情况下，卷积神经网络可以通过加深网络深度及宽度以提高网络性能。但当网络深度及宽度增加足够大时，网络的训练难度也越来越大，最终导致网络性能下降。经过对众多学者在相关领域的研究总结和思考，本发明对研究涉及的经典卷积神经网络模型进行两点改进。The application of convolutional neural networks in the field of image recognition continues to deepen, and excellent algorithms emerge in endlessly. In general, convolutional neural networks can improve network performance by deepening the depth and width of the network. However, when the depth and width of the network increase sufficiently, the training difficulty of the network becomes more and more difficult, which eventually leads to a decrease in network performance. After summarizing and thinking about the research of many scholars in related fields, the present invention makes two improvements to the classic convolutional neural network model involved in the research.
结合随机森林思想，从集成学习角度出发，对卷积神经网络常用模型进行改进。基于随机森林思想将多种常用卷积神经网络模型输出结果的平均值作为网络的最终输出结果。通过集成互不依赖的多种网络模型，以并行运算的方式，通过“投票法”得出输出结果。采用AlexNet、GoogLeNet、ResNet和DenseNet四种卷积神经网络常用模型建立集成学习规则，对上述四种卷积神经网络常用模型预测结果进行混合运算，取平均预测概率最大值对应货物类型为最终结果。建立的卷积神经网络常用模型集成学习规则如图4所示，改进后的模型命名为RF-CNN。Combined with the idea of random forest, from the perspective of ensemble learning, the commonly used models of convolutional neural networks are improved. Based on the random forest idea, the average value of the output results of various commonly used convolutional neural network models is used as the final output of the network. By integrating multiple network models that are not dependent on each other, the output results are obtained through the "voting method" in a parallel operation. Four commonly used convolutional neural network models, AlexNet, GoogLeNet, ResNet and DenseNet, are used to establish integrated learning rules, and the prediction results of the above four commonly used convolutional neural network models are mixed. The common model ensemble learning rules of the established convolutional neural network are shown in Figure 4, and the improved model is named RF-CNN.
为构建绿通车运载货物放射源图像识别模型，论文以Cifar10数据集对AlexNet、DenseNet、GoogleNet和ResNet-进行预训练，然后使用绿通车运载货物放射源图像进行重新训练以微调模型。这是图像识别训练常用的数据集。CIFAR-10数据集是由CIFAR(CandianInstitute For Advanced Research)收集整理的一个用于机器学习和图像识别问题的数据集。这个数据集共有60000张32*32的涵盖10个分类的彩色图片。In order to build the image recognition model of radioactive sources of goods carried by green traffic, the paper pre-trained AlexNet, DenseNet, GoogleNet and ResNet- with the Cifar10 dataset, and then used the images of radioactive sources of green traffic to carry out retraining to fine-tune the model. This is a commonly used dataset for image recognition training. The CIFAR-10 dataset is a dataset collected and organized by CIFAR (Candian Institute For Advanced Research) for machine learning and image recognition problems. This data set has a total of 60,000 32*32 color pictures covering 10 categories.
该数据集数量多、类别丰富，可满足绝大多数神经网络的训练。The data set has a large number and rich categories, which can meet the training of most neural networks.
3、绿通车运载货物放射源图像识别模型调优3. Optimizing the image recognition model of radioactive sources carried by green vehicles
在确定绿通车运载货物放射源图像识别模型框架基础上，需进一步优化模型性能。论文对绿通车运载货物放射源图像识别模型调优分为两步，首先提高数据集图像质量，然后使用数据迭代器解决内存占用过大问题，增加网络输入图像尺寸。On the basis of determining the framework of the image recognition model for radioactive sources of goods carried by green traffic, the performance of the model needs to be further optimized. In this paper, the optimization of the image recognition model of radioactive sources carried by green traffic is divided into two steps. First, improve the image quality of the dataset, and then use the data iterator to solve the problem of excessive memory usage and increase the size of the network input image.
网络将数据集中所有图像数据一次性加载到内存中，由于数据集图像数量达到上万余张，考虑到程序内存占用率过大将导致计算机内存不足(“内存爆炸”)，上述实验中输入图像尺寸仅采用32×32，经插值处理后分别转换为AlexNet、GoogLeNet、ResNet和DenseNet模型规定尺寸大小。由于输入图像尺寸过小，造成大量原始图像特征信息丢失，不利于网络性能的提升。The network loads all the image data in the data set into the memory at one time. Since the number of images in the data set reaches tens of thousands, considering that the excessive memory usage of the program will lead to insufficient computer memory ("memory explosion"), the input image size in the above experiment is Only 32×32 is used, and after interpolation processing, they are respectively converted into the specified sizes of AlexNet, GoogLeNet, ResNet and DenseNet models. Due to the small size of the input image, a large amount of original image feature information is lost, which is not conducive to the improvement of network performance.
为降低程序内存占用率，论文采用数据迭代器对程序进行优化，与之前的数据读取方式不同，迭代器按需读取数据，可有效解决内存爆炸问题。使用数据迭代器优化后的程序可应对大数据量读取问题。理论上，应尽可能增加图像尺寸以保证原始图像特征信息不丢失，当输入图像尺寸大于模型规定尺寸时，在进行卷积运算提取图像特征前，网络仍然会将图像尺寸降低至规定尺寸大小。因此论文将输入图像尺寸由32×32直接增加至AlexNet、GoogLeNet、ResNet和DenseNet模型规定尺寸大小，并分别设置实验，验证优化后的程序对网络性能的提升效果。In order to reduce the memory usage of the program, the paper uses a data iterator to optimize the program. Unlike the previous data reading method, the iterator reads data on demand, which can effectively solve the problem of memory explosion. The optimized program using the data iterator can deal with the problem of reading a large amount of data. In theory, the image size should be increased as much as possible to ensure that the original image feature information is not lost. When the input image size is larger than the specified size of the model, the network will still reduce the image size to the specified size before performing convolution operations to extract image features. Therefore, the paper directly increases the input image size from 32×32 to the specified size of AlexNet, GoogLeNet, ResNet and DenseNet models, and sets up experiments respectively to verify the network performance improvement effect of the optimized program.
4、验证实验结果4. Verify the experimental results
通过对模型的训练、验证及测试，并记录训练过程中损失函数及分类准确率的变化。损失函数能够反映模型对货物类型准确分类的能力。Through the training, verification and testing of the model, and record the changes in the loss function and classification accuracy during the training process. The loss function can reflect the ability of the model to accurately classify the cargo type.
Claims (7)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910683531.8ACN110390341B (en) | 2019-07-26 | 2019-07-26 | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910683531.8ACN110390341B (en) | 2019-07-26 | 2019-07-26 | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110390341A CN110390341A (en) | 2019-10-29 |
| CN110390341Btrue CN110390341B (en) | 2022-12-06 |
Family
ID=68287778
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910683531.8AActiveCN110390341B (en) | 2019-07-26 | 2019-07-26 | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110390341B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027456B (en)* | 2019-12-06 | 2023-06-20 | 四川杰森机电有限公司 | Mechanical water meter reading identification method based on image identification |
| CN111711815B (en)* | 2020-06-30 | 2021-09-28 | 福州大学 | Fast VVC intra-frame prediction method based on integrated learning and probability model |
| CN111860361B (en)* | 2020-07-24 | 2023-07-25 | 吉林大学 | Automatic identifier and identification method for green channel cargo scanning image entrainment |
| CN119106045B (en)* | 2024-11-08 | 2025-02-14 | 江西众加利高科技股份有限公司 | Green pass inspection method, device, equipment and storage medium for freight vehicles on highways |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145898A (en)* | 2017-04-14 | 2017-09-08 | 北京航星机器制造有限公司 | A Classification Method of Ray Image Based on Neural Network |
| WO2018028255A1 (en)* | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN109508650A (en)* | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
- 2019
- 2019-07-26CNCN201910683531.8Apatent/CN110390341B/enactiveActive
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018028255A1 (en)* | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN107145898A (en)* | 2017-04-14 | 2017-09-08 | 北京航星机器制造有限公司 | A Classification Method of Ray Image Based on Neural Network |
| CN109508650A (en)* | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于深度卷积神经网络的车型识别;石磊等;《计算机科学》;20180515(第05期);全文* |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110390341A (en) | 2019-10-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110390341B (en) | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods | |
| CN101713776B (en) | Neural network-based method for identifying and classifying visible components in urine | |
| CN110188774B (en) | A Classification and Recognition Method of Eddy Current Scanning Image Based on Deep Learning | |
| CN111914728B (en) | Hyperspectral remote sensing image semi-supervised classification method and device and storage medium | |
| CN111368896A (en) | A classification method of hyperspectral remote sensing images based on dense residual 3D convolutional neural network | |
| CN110232419A (en) | A kind of method of side slope rock category automatic identification | |
| Sungheetha et al. | Comparative study: statistical approach and deep learning method for automatic segmentation methods for lung CT image segmentation | |
| CN119338827B (en) | Surface detection method and system for precision fasteners | |
| CN108229551B (en) | Hyperspectral remote sensing image classification method based on compact dictionary sparse representation | |
| CN108764138A (en) | A kind of highlands cloud snow sorting technique cascading forest based on the more granularities of multidimensional | |
| CN108830243A (en) | Hyperspectral image classification method based on capsule network | |
| CN117474863A (en) | Chip surface defect detection method for compressed multi-head self-attention neural network | |
| CN111222545B (en) | Image classification method based on linear programming incremental learning | |
| CN106934418A (en) | A kind of insulator infrared diagnostics method based on convolution Recursive Networks | |
| CN113673556B (en) | A hyperspectral image classification method based on multi-scale dense convolutional network | |
| CN109191434A (en) | Image detecting system and detection method in a kind of cell differentiation | |
| CN115424006A (en) | Multi-source multi-level data fusion method applied to crop phenotypic parameter inversion | |
| CN118898610B (en) | A textile defect detection method and system based on machine vision | |
| CN118865142B (en) | Statistical analysis method for wild plant classified remote sensing image data | |
| TW201910929A (en) | Generating high resolution images from low resolution images for semiconductor applications | |
| CN116051382B (en) | A data enhancement method based on deep reinforcement learning generative adversarial neural network and super-resolution reconstruction | |
| CN103065296B (en) | High-resolution remote sensing image residential area extraction method based on edge feature | |
| CN119151936B (en) | Coal rock carbon content detection method and system based on multi-feature image spectroscopy technology | |
| CN118655099B (en) | Multi-dimensional hyperspectral image-based water quality parameter classification inversion method and system | |
| Hu et al. | Research on rice grain shape detection method based on machine vision |
Legal Events
| 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |