技术领域Technical Field
一种烟叶烘烤阶段智能识别方法，属于烟叶智能烘烤技术领域。The invention discloses an intelligent identification method for tobacco leaf baking stages, belonging to the technical field of tobacco leaf intelligent baking.
背景技术Background technique
智能算法作为人工智能的重要算法，在许多领域得到广泛应用，它表示输入与输出之间的非线性关系。机器视觉领域需要认识外界环境，以机器学习算法表示认识外界环境的过程。智能算法作为机器学习的重要分支，广泛应用于烟叶烘烤中的烘烤阶段判别。在烟叶烘烤中，不同的烟叶状态需要不同的温湿度控制工艺，利用智能判别算法建立烟叶图像与烟叶状态的非线性关系准确地识别烟叶所属阶段，有效地控制对应的温湿度。As an important algorithm of artificial intelligence, intelligent algorithm is widely used in many fields. It represents the nonlinear relationship between input and output. The field of machine vision requires the recognition of the external environment, and the process of recognizing the external environment is represented by machine learning algorithms. As an important branch of machine learning, intelligent algorithm is widely used in the identification of baking stages in tobacco leaf baking. In tobacco leaf baking, different tobacco leaf states require different temperature and humidity control processes. The intelligent discrimination algorithm is used to establish a nonlinear relationship between tobacco leaf images and tobacco leaf states to accurately identify the stage of tobacco leaves and effectively control the corresponding temperature and humidity.
烟叶智能判别算法是指在众多数据中寻找普适性特征，最直接的方式是通过迭代训练不断学习并更新参数表示烟叶图像与烟叶阶段之间的非线性关系。考虑到烟叶图像受到多种因素的影响，得到强泛化能力的非线性关系所花费的代价非常巨大。为了解决这一问题，通常稳定烟叶图像的成像环境和增加核函数增强非线性表示能力。稳定烟叶图像的成像环境主要包括固定的光源及固定的摄像机安装位置。固定的光源能够减少光照变化对图像色彩的影响，保证同一阶段所表现的色彩信息相近，固定的摄像机安装位置减少摄像机角度对烟叶形状的影响，通过稳定成像环境降低环境复杂度；增加核函数能够增强特征表示能力，更有效的描述烟叶图像与烟叶阶段之间的关系，在降低环境复杂度的基础上，增强特征表示能力，很大程度上增强特征表示的普适性。The tobacco leaf intelligent discrimination algorithm refers to finding universal features in a large number of data. The most direct way is to continuously learn and update parameters through iterative training to represent the nonlinear relationship between tobacco leaf images and tobacco leaf stages. Considering that tobacco leaf images are affected by many factors, the cost of obtaining a nonlinear relationship with strong generalization ability is very huge. In order to solve this problem, the imaging environment of tobacco leaf images is usually stabilized and the kernel function is added to enhance the nonlinear representation ability. The imaging environment of the stable tobacco leaf image mainly includes a fixed light source and a fixed camera installation position. The fixed light source can reduce the impact of illumination changes on the color of the image, ensure that the color information shown in the same stage is similar, and the fixed camera installation position reduces the impact of the camera angle on the shape of the tobacco leaf, and reduces the complexity of the environment by stabilizing the imaging environment; adding the kernel function can enhance the feature representation ability, more effectively describe the relationship between tobacco leaf images and tobacco leaf stages, and enhance the feature representation ability on the basis of reducing the complexity of the environment, which greatly enhances the universality of the feature representation.
由于烟叶阶段表示的复杂性和烟叶图像规模的不断增加，直接构建烟叶图像与烟叶阶段之间非线性关系越来越困难，普适性特征也越来越难以搜索。现有的工作忽视了智能算法之间的融合，仅仅以构建非线性关系或数据统计单独进行阶段的判别，造成算法执行过程复杂度增加或效果稳定性不佳。所以提出一种新的智能算法兼具非线性表示的稳定性、泛化性及低复杂性是十分具有挑战性且富有实际意义的。Due to the complexity of tobacco leaf stage representation and the increasing size of tobacco leaf images, it is increasingly difficult to directly construct the nonlinear relationship between tobacco leaf images and tobacco leaf stages, and it is increasingly difficult to search for universal features. Existing work ignores the integration of intelligent algorithms and only distinguishes stages by constructing nonlinear relationships or data statistics, which increases the complexity of the algorithm execution process or causes poor stability of the effect. Therefore, it is very challenging and meaningful to propose a new intelligent algorithm that combines the stability, generalization and low complexity of nonlinear representation.
发明内容Summary of the invention
本发明要解决的技术问题是：克服现有技术的不足，提供一种使用网络二分类和数据统计实现烟叶阶段识别，降低算法的复杂度，提高效果的稳定性的烟叶烘烤阶段智能识别方法。The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for intelligently identifying tobacco leaf baking stages by using network binary classification and data statistics to realize tobacco leaf stage identification, reduce the complexity of the algorithm, and improve the stability of the effect.
本发明解决其技术问题所采用的技术方案是：该烟叶烘烤阶段智能识别方法，其特征在于：包括如下步骤：The technical solution adopted by the present invention to solve the technical problem is: the tobacco leaf baking stage intelligent identification method is characterized by comprising the following steps:
S1建立模型，去除烟叶之外的背景部分，并提取烟叶区域；S1 builds a model, removes the background outside the tobacco leaves, and extracts the tobacco leaf area;
S2统计烟叶区域内颜色信息，寻找不同阶段特有的颜色信息特征；S2 counts the color information in the tobacco leaf area and searches for the color information features specific to different stages;
S3完成烟叶颜色信息的提取，分析各阶段颜色信息的特点；S3 completes the extraction of tobacco leaf color information and analyzes the characteristics of color information at each stage;
S4判断完成烟叶阶段后，将烟叶阶段转换为烟叶烘烤工艺阶段。After S4 determines that the tobacco leaf stage is completed, the tobacco leaf stage is converted into a tobacco leaf baking process stage.
优选的，所述的模型为：Preferably, the model is:
Y＝Fω(X)；Y =Fω (X);
其中，X表示烟叶图像，Y表示烟叶二值图像，Fω关于权重ω的非线性关系式。Where X represents the tobacco leaf image, Y represents the tobacco leaf binary image, andFω is the nonlinear relationship between the weight ω.
优选的，通过前向传播及反向误差传播来求得关于参数ω的非线性关系Fω。Preferably, the nonlinear relationship Fω with respect to the parameter ω is obtained through forward propagation and backward error propagation.
优选的，所述的前向传播依次包括初始特征提取阶段、多支路特征提取阶段以及特征融合阶段并进行上采样获取烟叶区域。Preferably, the forward propagation includes an initial feature extraction stage, a multi-branch feature extraction stage and a feature fusion stage in sequence, and upsampling is performed to obtain the tobacco leaf region.
优选的，在初始特征提取阶段，下一隐藏层的输出为：Preferably, in the initial feature extraction stage, the output of the next hidden layer is:
Yl+1＝σ(ωl+1Yl+Bl+1)；Yl+1 =σ(ωl+1 Yl +Bl+1 );
其中，l表示当前隐藏层，多维矩阵Yl为当前隐藏层的输出结果，ωl+1为下一隐藏层的权值参数，Bl+1为下一隐藏层的偏置参数，σ是非线性的激活函数；Where l represents the current hidden layer, the multidimensional matrix Yl is the output result of the current hidden layer, ωl+1 is the weight parameter of the next hidden layer, Bl+1 is the bias parameter of the next hidden layer, and σ is a nonlinear activation function;
非线性的激活函数σ为：The nonlinear activation function σ is:
优选的，所述的多支路特征提取阶段的多支路特征为：Preferably, the multi-branch features in the multi-branch feature extraction stage are:
其中，为高分辨率融合特征，/>为中分辨率初始特征，/>为低分辨率初始特征，/>为卷积操作，ωH、、ωM、ωL分别指高分辨率模板卷积权重、中分辨率模板卷积权重、低分辨率模板卷积权重，U()指下采样操作。in, is a high-resolution fusion feature,/> is the medium resolution initial feature,/> is the low-resolution initial feature,/> is a convolution operation, ωH , ωM , ωL refer to the high-resolution template convolution weight, the medium-resolution template convolution weight, and the low-resolution template convolution weight, respectively, and U() refers to the downsampling operation.
优选的，所述方法还包括，烟叶图像X和提取烟叶区域Y之间的关系为：Preferably, the method further includes that the relationship between the tobacco leaf image X and the extracted tobacco leaf region Y is:
其中，YF为低、中、高分辨率融合特征。Among them, YF is the low, medium and high resolution fusion features.
优选的，所述方法还包括，利用交叉熵J描述误差的大小，其中交叉熵J为：Preferably, the method further comprises using cross entropy J to describe the magnitude of the error, wherein the cross entropy J is:
其中，N、n、c、yij、pij分别指指示函数值、结果预测概率值；Among them, N, n, c, yij , and pij refer to the indicator function value and the result prediction probability value respectively;
其中，Yij、Tij分别指结果预测值、真实值。Among them,Yij andTij refer to the predicted value and the true value of the result respectively.
优选的，所述方法还包括，前一层的误差大小δl-1为：Preferably, the method further includes that the error size δl-1 of the previous layer is:
δl-1＝ωlδlσ′(Yl-1)；δl-1 =ωl δl σ′(Yl-1 );
其中，δl为第l层的误差，ωl、σ′分别指第l层权重、激活函数的一阶导数；Among them, δl is the error of the lth layer, ωl and σ′ refer to the first-order derivatives of the weight and activation functions of the lth layer respectively;
第l层所分担误差J的比重以梯度为：The proportion of error J shared by the lth layer is based on the gradient:
l-1的权值参数：The weight parameter of l-1 is:
其中，η分别指第l-1层更新的权重、衰减权重。in, η refers to the updated weight and decay weight of the l-1th layer respectively.
优选的，所述方法还包括，黄色面积比ratio_Y和绿色的面积比ration_G分别为：Preferably, the method further includes that the yellow area ratio ratio_Y and the green area ratio ration_G are respectively:
其中，Ri，j、Gi，j、Bi，j分别为任意位置(i，j)红、绿、蓝的颜色值，如果与黄色相接近，此点划分为黄色并记为V(i，j)＝[1，0]，如果与绿色相接近，此点划分为绿色并记为V(i，j)＝[1，0]。Among them, Ri,j , Gi,j , and Bi,j are the red, green, and blue color values of any position (i,j), respectively. If it is close to yellow, this point is classified as yellow and recorded as V(i,j) = [1, 0]; if it is close to green, this point is classified as green and recorded as V(i,j) = [1, 0].
与现有技术相比，本发明所具有的有益效果是：Compared with the prior art, the present invention has the following beneficial effects:
本烟叶烘烤阶段智能识别方法采用智能算法融合的方式，建立的烟叶阶段判别算法获得更高的识别准确率和更强的泛化能力；采用颜色特征进行烟叶阶段识别，符合实际烘烤过程烟叶状态变化规律，缓解烘烤阶段反复跳跃及不稳定现象；网络结构完成复杂性低的二分类任务，检测烟叶区域，增加了检测稳定性及陌生环境的适应性。The tobacco leaf baking stage intelligent recognition method adopts the method of intelligent algorithm fusion to establish a tobacco leaf stage discrimination algorithm with higher recognition accuracy and stronger generalization ability; the tobacco leaf stage recognition is carried out by using color features, which conforms to the law of tobacco leaf state change in the actual baking process and alleviates the repeated jumps and instability in the baking stage; the network structure completes the low-complexity binary classification task and detects the tobacco leaf area, which increases the detection stability and adaptability to unfamiliar environments.
本发明实现判别烟叶阶段分为两个阶段进行。第一阶段，网络二分类，即去除背景并提取烟叶部分，在该阶段，各个阶段的烟叶区域被准确的提取。第二阶段，对烟叶区域内色彩信息进行数据统计，挖掘烟叶不同阶段的特征信息，在该阶段，利用烟叶区域内特征信息判别烟叶所处阶段。本发明以第一阶段为基础，解决烟叶区域提取的泛化能力问题，重点为网络结构表示的语义特征的普适性表示。第二阶段以第一阶段结果为基础，挖掘烟叶区域内的特征信息，重点寻找烟叶不同阶段特有的特征信息，使用颜色区别不同阶段并快速判断烟叶所处阶段。The present invention realizes the identification of tobacco leaf stages in two stages. In the first stage, the network is classified into two categories, that is, the background is removed and the tobacco leaf part is extracted. In this stage, the tobacco leaf areas of each stage are accurately extracted. In the second stage, data statistics are performed on the color information in the tobacco leaf area, and the characteristic information of the tobacco leaves at different stages is mined. In this stage, the characteristic information in the tobacco leaf area is used to identify the stage of the tobacco leaves. Based on the first stage, the present invention solves the problem of the generalization ability of tobacco leaf area extraction, focusing on the universal representation of the semantic features represented by the network structure. The second stage is based on the results of the first stage, mining the characteristic information in the tobacco leaf area, focusing on finding the characteristic information unique to the different stages of the tobacco leaves, using color to distinguish the different stages and quickly determine the stage of the tobacco leaves.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1为烟叶区域提取模型的流程图。FIG1 is a flow chart of the tobacco leaf region extraction model.
图2为烟叶阶段识别流程图。Figure 2 is a flowchart of tobacco leaf stage identification.
图3为烟叶工艺数字化示意图。Figure 3 is a schematic diagram of the digitalization of tobacco leaf processing.
图4为烟叶烘烤阶段智能识别流程图。Figure 4 is a flowchart of intelligent identification during the tobacco leaf baking stage.
具体实施方式Detailed ways
下面结合具体实施例对本发明做进一步说明，然而熟悉本领域的人们应当了解，在这里结合附图给出的详细说明是为了更好的解释，本发明的结构必然超出了有限的这些实施例，而对于一些等同替换方案或常见手段，本文不再做详细叙述，但仍属于本申请的保护范围。The present invention is further described below in conjunction with specific embodiments. However, people familiar with the art should understand that the detailed description given here in conjunction with the drawings is for better explanation, and the structure of the present invention necessarily exceeds these limited embodiments. For some equivalent replacement schemes or common means, they are no longer described in detail herein, but still belong to the scope of protection of the present application.
图1～4是本发明的最佳实施例，下面结合附图1～4对本发明做进一步说明。1 to 4 are the best embodiments of the present invention, and the present invention will be further described below in conjunction with FIGS. 1 to 4 .
如图1～4所示：一种烟叶烘烤阶段智能识别方法，包括如下步骤：As shown in Figures 1 to 4: A method for intelligently identifying tobacco leaf baking stages comprises the following steps:
S1建立模型，去除烟叶之外的背景部分，并提取烟叶区域；S1 builds a model, removes the background outside the tobacco leaves, and extracts the tobacco leaf area;
S2统计烟叶区域内颜色信息，寻找不同阶段特有的颜色信息特征；S2 counts the color information in the tobacco leaf area and searches for the color information features specific to different stages;
S3完成烟叶颜色信息的提取，分析各阶段颜色信息的特点；S3 completes the extraction of tobacco leaf color information and analyzes the characteristics of color information at each stage;
S4判断完成烟叶阶段后，将烟叶阶段转换为烟叶烘烤工艺阶段。After S4 determines that the tobacco leaf stage is completed, the tobacco leaf stage is converted into a tobacco leaf baking process stage.
烟叶烘烤阶段智能识别包含四个过程：建立模型并提取烟叶区域；统计烟叶区域内颜色信息，寻找不同阶段特有的颜色信息特征；烟叶阶段判别；输出烟叶阶段。首先，建立并训练网络模型，建立烟叶图像与烟叶区域之间的非线性关系。输入烟叶图像，通过训练完成的网络模型将烟叶区域从背景中提取出来；随后统计烟叶区域内黄色及绿色像素值的面积、面积占比等；根据不同烘烤阶段黄绿色面积占比不同的特点，计算出区分阶段间黄色或绿色面积占比的阈值，并对烟叶图像进行阶段判别；最后将判别的阶段。Intelligent identification of tobacco leaf baking stages includes four processes: building a model and extracting tobacco leaf areas; counting the color information in the tobacco leaf area and finding the color information characteristics unique to different stages; distinguishing tobacco leaf stages; and outputting tobacco leaf stages. First, a network model is built and trained to establish a nonlinear relationship between tobacco leaf images and tobacco leaf areas. Input tobacco leaf images, and extract tobacco leaf areas from the background through the trained network model; then count the area and area proportion of yellow and green pixel values in the tobacco leaf area; according to the different characteristics of the yellow-green area proportion in different baking stages, calculate the threshold for distinguishing the yellow or green area proportion between stages, and distinguish the tobacco leaf images; finally, the distinguished stage.
具体的，建立模型并提取烟叶区域的具体方法如下：Specifically, the specific method of establishing the model and extracting the tobacco leaf area is as follows:
烟叶区域提取算法用来去除烟叶之外的背景部分。输入烟叶图像X，使用Y＝Fω(X)表示烟叶图像X与提取烟叶区域Y之间的非线性关系。The tobacco leaf region extraction algorithm is used to remove the background outside the tobacco leaf. Input the tobacco leaf image X, and use Y=Fω (X) to represent the nonlinear relationship between the tobacco leaf image X and the extracted tobacco leaf region Y.
求得关于参数ω的非线性关系Fω需要两个步骤：前向传播及反向误差传播。Finding the nonlinear relationshipFω about the parameter ω requires two steps: forward propagation and backward error propagation.
前向传播的具体过程为：具有多隐藏层的神经网络结构，分为三部分。第一部分初始特征提取阶段；第二部分多支路特征提取阶段；第三部分特征融合阶段并进行上采样获取烟叶区域。初始特征提取阶段，l表示当前隐藏层，多维矩阵Yl为当前隐藏层的输出结果，ωl+1为下一隐藏层的权值参数，Bl+1为下一隐藏层的偏置参数，计算得到下一隐藏层的输出为：The specific process of forward propagation is as follows: the neural network structure with multiple hidden layers is divided into three parts. The first part is the initial feature extraction stage; the second part is the multi-branch feature extraction stage; the third part is the feature fusion stage and upsampling to obtain the tobacco leaf area. In the initial feature extraction stage, l represents the current hidden layer, the multidimensional matrix Yl is the output result of the current hidden layer, ωl+1 is the weight parameter of the next hidden layer, Bl+1 is the bias parameter of the next hidden layer, and the output of the next hidden layer is calculated as:
Yl+1＝σ(ωl+1Yl+Bl+1)；Yl+1 =σ(ωl+1 Yl +Bl+1 );
σ是非线性的激活函数；σ is a nonlinear activation function;
输入图像X经过多个隐藏层，得到初始特征Yini。多支路特征提取，将初始特征进行不同程度的下采样得到多维矩阵三个低分辨率、中分辨率、高分辨率特征图，分别输入三个分支得到多支路特征。The input image X passes through multiple hidden layers to obtain the initial featureYini . Multi-branch feature extraction, the initial features are downsampled to different degrees to obtain a multi-dimensional matrix Three low-resolution, medium-resolution, and high-resolution feature maps are input into three branches respectively to obtain multi-branch features.
其中，为高分辨率融合特征，/>为中分辨率初始特征，/>为低分辨率初始特征，/>为卷积操作，ωH、、ωM、ωL分别指高分辨率模板卷积权重、中分辨率模板卷积权重、低分辨率模板卷积权重，U()指下采样操作。in, is a high-resolution fusion feature,/> is the medium resolution initial feature,/> is the low-resolution initial feature,/> is a convolution operation, ωH , ωM , ωL refer to the high-resolution template convolution weight, the medium-resolution template convolution weight, and the low-resolution template convolution weight, respectively, and U() refers to the downsampling operation.
多支路特征为具有不同分辨率大小的多维矩阵，随后进行多支路的特征融合，低分辨率与中分辨率特征进行卷积操作并上采样，高分辨率特征进行卷积操作。The multi-branch features are multi-dimensional matrices with different resolution sizes, and then the multi-branch features are fused, the low-resolution and medium-resolution features are convolved and upsampled, and the high-resolution features are convolved.
其中，YF为低、中、高分辨率融合特征。Among them, YF is the low, medium and high resolution fusion features.
输入烟叶图像X经过初始特征提取阶段、多支路特征提取阶段和特征融合阶段，建立了烟叶图像X和提取烟叶区域Y之间的关系。The input tobacco leaf image X undergoes the initial feature extraction stage, the multi-branch feature extraction stage and the feature fusion stage, and the relationship between the tobacco leaf image X and the extracted tobacco leaf region Y is established.
反向误差传播过程为：通过前向传播建立的烟叶图像X和提取烟叶区域Y之间的关系存在较大的误差，为使得建立的关系更加符合实际情况，需要对其进行调整。经过计算提取的烟叶区域Y与实际的烟叶区域T存在误差，利用交叉熵J描述误差的大小。The reverse error propagation process is as follows: there is a large error in the relationship between the tobacco leaf image X and the extracted tobacco leaf area Y established by forward propagation. In order to make the established relationship more in line with the actual situation, it needs to be adjusted. After calculation, there is an error between the extracted tobacco leaf area Y and the actual tobacco leaf area T. The cross entropy J is used to describe the size of the error.
其中，N、n、c、yij、pij分别指指示函数值、结果预测概率值；Among them, N, n, c, yij , and pij refer to the indicator function value and the result prediction probability value respectively;
其中，Yij是指示函数，作为二分类网络，指示函数可取值为0和1。Among them,Yij is the indicator function. As a binary classification network, the indicator function can take values 0 and 1.
其中，Yij、TijTij分别指结果预测值、真实值。Among them,Yijand Tijrefer to the predicted value and the true value of the result respectively.
随后，将误差值J传递到每个隐藏层并计算每个隐藏层分担误差J的比重。第l层的误差为δl，则其前一层的误差大小为δl-1。Then, the error value J is passed to each hidden layer and the proportion of each hidden layer sharing the error J is calculated. If the error of the lth layer is δl , then the error of the previous layer is δl-1 .
δl-1＝ωlδlσ′(Yl-1)；δl-1 =ωl δl σ′(Yl-1 );
第l层所分担误差J的比重以梯度的链式法则计算得到。The proportion of error J shared by the lth layer is calculated using the chain rule of gradient.
最后更新l-1的权值参数：Finally update the weight parameters of l-1:
通过前向传播计算误差，反向误差传播对误差进行修正，通过多次迭代学习到较优的烟叶区域Y与实际的烟叶区域T之间的非线性关系，实现烟叶区域的提取。The error is calculated through forward propagation and corrected through backward error propagation. The nonlinear relationship between the optimal tobacco leaf area Y and the actual tobacco leaf area T is learned through multiple iterations to achieve the extraction of the tobacco leaf area.
统计烟叶区域内颜色信息，寻找不同阶段特有的颜色信息特征，具体过程为：Count the color information in the tobacco leaf area and find the color information characteristics unique to different stages. The specific process is as follows:
去除背景之后，图片仅剩烟叶区域。烘烤过程中烟叶由绿变黄，所以黄、绿颜色变化是其重要的特征。Ri，j、Gi，j、Bi，j分别为任意位置(i，j)红、绿和蓝的颜色值，如果与黄色相接近，此点划分为黄色并记为V(i，j)＝[1，0]，如果与绿色相接近，此点划分为绿色并记为V(i，j)＝[1，0]。统计黄色与绿色的面积比，分别记为ratio_Y和ratio_G。After removing the background, only the tobacco leaf area remains in the image. During the baking process, tobacco leaves change from green to yellow, so the change in yellow and green color is its important feature.R i,j , Gi,j , and Bi,j are the red, green, and blue color values of any position (i, j), respectively. If it is close to yellow, this point is classified as yellow and recorded as V(i, j) = [1, 0]. If it is close to green, this point is classified as green and recorded as V(i, j) = [1, 0]. The area ratios of yellow and green are counted and recorded as ratio_Y and ratio_G, respectively.
烟叶阶段判别的具体过程为：The specific process of tobacco leaf stage identification is as follows:
完成了烟叶颜色信息的提取，分析各阶段颜色信息的特点。输入黄色面积占比ratio_Y和绿色面积占比ratio_G，当黄色面积占比小不大于0.35时，输出[1 0 0 0 0 0 00 0 0]，即判断为第一阶段；当黄色面积占比大于0.35且不大于0.75时，输出[0 1 0 0 0 00 0 0 0 0]，即判断为第二阶段；当黄色面积占比介于0.75与0.86之间时，输出[0 0 1 0 00 0 0 0 0]，即判断为第三阶段；当黄色面积占比介于0.86与0.92之间时，输出[0 0 0 1 00 0 0 0 0]，即判断为第四阶段；前四个阶段为变黄期，主要通过黄色面积进行阶段判定。接下来时定色期，根据定色期特点，使用绿色面积对其进行判断。当绿色面积占比不大于0.04时，输出[0 0 0 0 1 0 0 0 0 0]，即判断为第五阶段；当绿色面积介于0.04和0.025之间时，输出[0 0 0 0 0 1 0 0 0 0]，即判断为第六阶段；当绿色面积介于0.025和0.015之间时，输出[0 0 0 0 0 0 1 0 0 0]，即判断为第七阶段。干筋期，主要为干燥剩余水分，黄绿色无明显变换，以干燥时间进行阶段判定。进入干筋期时，初始化时间为0，当时间小于14小时，输出[0 0 0 0 0 0 0 1 0 0]，即判断为第八阶段；当时间介于14与28小时，输出[0 00 0 0 0 0 0 1 0]，即判断为第九阶段；当时间介于28与52小时，输出[0 0 0 0 0 0 0 0 01]，即判断为第十阶段；时间超过52小时，烘烤停止。The extraction of tobacco leaf color information is completed, and the characteristics of color information in each stage are analyzed. Input the yellow area ratio_Y and the green area ratio ratio_G. When the yellow area ratio is less than or equal to 0.35, output [1 0 0 0 0 0 0 0 0], which is judged as the first stage; when the yellow area ratio is greater than 0.35 and less than 0.75, output [0 1 0 0 0 00 0 0 0], which is judged as the second stage; when the yellow area ratio is between 0.75 and 0.86, output [0 0 1 0 00 0 0 0 0], which is judged as the third stage; when the yellow area ratio is between 0.86 and 0.92, output [0 0 0 1 00 0 0 0 0], which is judged as the fourth stage; the first four stages are yellowing stages, and the stage judgment is mainly based on the yellow area. The next is the color fixing stage. According to the characteristics of the color fixing stage, the green area is used to judge it. When the green area ratio is no more than 0.04, the output is [0 0 0 0 1 0 0 0 0 0], which is judged as the fifth stage; when the green area is between 0.04 and 0.025, the output is [0 0 0 0 0 1 0 0 0 0], which is judged as the sixth stage; when the green area is between 0.025 and 0.015, the output is [0 0 0 0 0 0 1 0 0 0], which is judged as the seventh stage. In the dry tendon stage, it is mainly the residual moisture after drying, and there is no obvious change in the yellow-green color. The stage is judged by the drying time. When entering the drying period, the initialization time is 0. When the time is less than 14 hours, the output is [0 0 0 0 0 0 0 1 0 0], which is judged as the eighth stage. When the time is between 14 and 28 hours, the output is [0 00 0 0 0 0 0 1 0], which is judged as the ninth stage. When the time is between 28 and 52 hours, the output is [0 0 0 0 0 0 0 0 01], which is judged as the tenth stage. When the time exceeds 52 hours, baking stops.
输出烟叶阶段的具体过程为：The specific process of tobacco leaf output stage is as follows:
判断完成烟叶阶段后，将烟叶阶段转换为烟叶烘烤工艺阶段。烟叶工艺数字化算法是烟叶烘烤工艺与计算机能够识别语言相互转换的算法。烟叶烘烤的十个阶段，分别对应1、2、3、…、10个数字表示，转换为计算机能够识别的语言输出为10个10维的0、1的向量，向量中值1所在位置的索引即表示对应的烘烤阶段。这部分主要用于烟叶烘烤工艺与计算机语言的相互转换。After the tobacco leaf stage is determined to be completed, the tobacco leaf stage is converted into the tobacco leaf baking process stage. The tobacco leaf process digitization algorithm is an algorithm for converting tobacco leaf baking process into a language that can be recognized by computers. The ten stages of tobacco leaf baking correspond to 1, 2, 3, ..., 10 digital representations, which are converted into a language that can be recognized by computers and output as 10 10-dimensional vectors of 0 and 1. The index of the position of the value 1 in the vector represents the corresponding baking stage. This part is mainly used for the conversion between tobacco leaf baking process and computer language.
以上所述，仅是本发明的较佳实施例而已，并非是对本发明作其它形式的限制，任何熟悉本专业的技术人员可能利用上述揭示的技术内容加以变更或改型为等同变化的等效实施例。但是凡是未脱离本发明技术方案内容，依据本发明的技术实质对以上实施例所作的任何简单修改、等同变化与改型，仍属于本发明技术方案的保护范围。The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310211945.7ACN116434045B (en) | 2023-03-07 | 2023-03-07 | Intelligent identification method for tobacco leaf baking stages |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310211945.7ACN116434045B (en) | 2023-03-07 | 2023-03-07 | Intelligent identification method for tobacco leaf baking stages |
| Publication Number | Publication Date |
|---|---|
| CN116434045A CN116434045A (en) | 2023-07-14 |
| CN116434045Btrue CN116434045B (en) | 2024-06-14 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310211945.7AActiveCN116434045B (en) | 2023-03-07 | 2023-03-07 | Intelligent identification method for tobacco leaf baking stages |
| Country | Link |
|---|---|
| CN (1) | CN116434045B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117457101B (en)* | 2023-12-22 | 2024-03-26 | 中国农业科学院烟草研究所(中国烟草总公司青州烟草研究所) | A method, medium and system for predicting moisture content of cured tobacco leaves |
| CN118411368B (en)* | 2024-07-04 | 2024-09-13 | 中国农业科学院烟草研究所(中国烟草总公司青州烟草研究所) | Intelligent detection method, medium and system for baked tobacco leaves |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113919442A (en)* | 2021-02-24 | 2022-01-11 | 北京优创新港科技股份有限公司 | Tobacco maturity state recognition model based on convolutional neural network |
| CN114304699A (en)* | 2021-12-27 | 2022-04-12 | 上海创和亿电子科技发展有限公司 | Method for extracting green and yellow ratio of primary flue-cured tobacco leaves |
| CN114943876A (en)* | 2022-06-20 | 2022-08-26 | 南京信息工程大学 | Cloud and cloud shadow detection method and device for multi-level semantic fusion and storage medium |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103919258A (en)* | 2013-03-02 | 2014-07-16 | 重庆大学 | Densification tobacco flue-cure dry-wet bulb temperature automatic control technique based on tobacco image processing |
| CN109872328B (en)* | 2019-01-25 | 2021-05-07 | 腾讯科技（深圳）有限公司 | Brain image segmentation method, device and storage medium |
| CN112949663B (en)* | 2021-01-28 | 2024-01-23 | 贵州省烟草科学研究院 | Tobacco leaf baking control system and method based on image recognition |
| CN112818827B (en)* | 2021-01-28 | 2024-03-08 | 贵州省烟草科学研究院 | Method for judging stage temperature control point in tobacco leaf baking process based on image recognition |
| CN113080500B (en)* | 2021-04-01 | 2022-10-28 | 深圳九章聚源人工智能科技有限公司 | Tobacco leaf baking method and system based on neural network and related components thereof |
| CN113643304B (en)* | 2021-08-05 | 2024-09-06 | 应急管理部天津消防研究所 | Real-time smoke segmentation method integrating multi-resolution characterization |
| CN113688808B (en)* | 2021-10-26 | 2022-02-11 | 南京信息工程大学 | Landslide mass identification method based on Laplacian pyramid remote sensing image fusion |
| CN115115549B (en)* | 2022-07-01 | 2024-12-13 | 安徽理工大学 | Image enhancement model, method, device and storage medium with multi-branch fusion attention mechanism |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113919442A (en)* | 2021-02-24 | 2022-01-11 | 北京优创新港科技股份有限公司 | Tobacco maturity state recognition model based on convolutional neural network |
| CN114304699A (en)* | 2021-12-27 | 2022-04-12 | 上海创和亿电子科技发展有限公司 | Method for extracting green and yellow ratio of primary flue-cured tobacco leaves |
| CN114943876A (en)* | 2022-06-20 | 2022-08-26 | 南京信息工程大学 | Cloud and cloud shadow detection method and device for multi-level semantic fusion and storage medium |
| Publication number | Publication date |
|---|---|
| CN116434045A (en) | 2023-07-14 |
| Publication | Publication Date | Title |
|---|---|---|
| CN110276765B (en) | Image panorama segmentation method based on multi-task learning deep neural network | |
| CN108830855B (en) | Full convolution network semantic segmentation method based on multi-scale low-level feature fusion | |
| Sahin et al. | Yolodrone: Improved yolo architecture for object detection in drone images | |
| CN109614996B (en) | Weak visible light and infrared image fusion identification method based on generation countermeasure network | |
| CN109949317B (en) | Semi-supervised image example segmentation method based on gradual confrontation learning | |
| CN108960127B (en) | Re-identification of occluded pedestrians based on adaptive deep metric learning | |
| CN109344701B (en) | Kinect-based dynamic gesture recognition method | |
| CN108399428B (en) | Triple loss function design method based on trace ratio criterion | |
| CN116434045B (en) | Intelligent identification method for tobacco leaf baking stages | |
| CN111768432A (en) | Moving object segmentation method and system based on Siamese deep neural network | |
| CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
| CN112365462B (en) | An Image-Based Change Detection Method | |
| CN110991349B (en) | A Lightweight Vehicle Attribute Recognition Method Based on Metric Learning | |
| CN112232151A (en) | An Iterative Aggregation Neural Network High Score Remote Sensing Scene Classification Method with Embedding Attention Mechanism | |
| CN113128308B (en) | Pedestrian detection method, device, equipment and medium in port scene | |
| CN111046789A (en) | Pedestrian re-identification method | |
| Zhang et al. | Joint distribution alignment via adversarial learning for domain adaptive object detection | |
| CN113361466A (en) | Multi-modal cross-directed learning-based multi-spectral target detection method | |
| CN110929099A (en) | Short video frame semantic extraction method and system based on multitask learning | |
| Hu et al. | Semantic segmentation based on semantic edge optimization | |
| CN111783673A (en) | An Improved Method of Video Segmentation Based on OSVOS | |
| CN111680602A (en) | Pedestrian re-identification method and model architecture based on dual-stream hierarchical feature correction | |
| CN111126173A (en) | A high-precision face detection method | |
| CN113128460B (en) | Multi-resolution person re-identification method based on knowledge distillation | |
| CN116503618A (en) | A salient target detection method and device based on multi-modal multi-level feature aggregation |
| 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 |