Vision-Based Classification And Recognition Of Cigarettes On Transmission Platform

With the rapid development of the tobacco industry,the number of business orders is on the increase.In order to improve the efficiency of logistics management,many tobacco companies have adopted automated or semi-automated sorting and distribution systems of cigarettes.There are the problems of lack of accuracy in the existing systems of cigarette sorting.If the order is wrong,it will not only affect the economic interests of tobacco companies and customers,but also seriously affect the reputation and image of the companies.It is difficult to check the sorting results of cigarettes by manual method,so the technology of cigarettes accuracy checking has become an important research direction.By using of image acquisition device to obtain transmission platform video data,it is of great research significance to realize the automatic check of cigarettes rod orders by image classification and recognition technology.For sorting delivery platform,this paper designs cigarettes classification and identification algorithm.The main research contents are as follows:(1)This paper studies the theory of image preprocessing,including RGB color space,color image grayscale,image morphological processing theory.This paper counts the color distribution characteristics of smoke-free transmission platform image.It is conducive to more accurate distinction between cigarette and non-cigarette area,according to the distribution characteristics of pixel-by-pixel non-uniform light correction.(2)This paper studies threshold-based image segmentation,edge-based image segmentation and region-based image segmentation algorithms.The background of the transmission platform is similar to the pixel,the pixel between the cigarette strips is changeable,but all have the characteristics that are different from the background pixel.Based on this,the background differential threshold segmentation is used to distinguish between non-smoking area and smoke area.(3)This paper studies the theory of deep learning.In this paper,we introduce a common model of deep neural network,including deep belief network,recurrent neural network and convolution neural network.The basic structure of convolution neural network is introduced in detail,including convolution layer,activation function,pooling layer,loss function.Finally,back propagation method for neural network learning is deduced.(4)This paper studies the contour segmentation method and uses the background difference threshold segmentation to obtain binary images,using the morphology processing to improve the segmentation integrity.In this paper,the approximate polygon curve of connected domain is extracted,and a series of contour inflection points are obtained.According to the shape,size and position of cigarettes,fitting the outline inflection point,we get four vertex information of cigarettes.If multiple cigarettes are connected together,the foreground area is further divided according to the moving direction of the conveyor belt and the historical location information of the cigarettes segment.By counting the results of segmentation between video frames,the number of cigarettes passed is confirmed.The recall rate of the test data is 99.75% and the accuracy rate is 100%.(5)This paper studies the image classification model in deep learning and uses the convolutional neural network to classify cigarettes rod.According to class sorting image segmentation of the cigarettes data,the model is trained by two methods.The first method is to train the parameters of the model network directly with the cigarettes data.The second method is to use the open data set to pre-train the network model and then use the cigarettes data to carry out the migration learning and fine-tune the network parameters.In this paper,we compared the two training methods to get the model accuracy,compared the different learning rates on the model convergence,and 100% accuracy was obtained on the verification set,which verified the performance of the algorithm.