Research On Shelf Commodity Detection Algorithm Based On Convolutional Neural Network

With the development of science and technology,intelligent has gradually become a trend.The emergence of more and more intelligent unmanned supermarkets and intelligent unmanned containers have made artificial intelligence technology successfully implemented.However,there are also many traditional supermarkets that rely on manual inventory methods to manage commodities,which costs a lot of manpower and time.Therefore,we mainly study the application of deep convolutional neural network in the detection of supermarket shelf commodities,so as to realize the intelligent management of supermarket commodities.To solve the problem that there is no open-source commodity dataset for research,this paper constructs a relatively complete commodity dataset by self-collection and data enhancement,and then manually labels the target area.By reading the literature,YOLOv3 algorithm based on regression performs very well in both speed and precision.So YOLOv3 network is used for the commodity detection.During the commodity detection,it is found that some of the commodities which occupy a large proportion and those that have partial occlusion are missed.In view of this problem,it is improved and optimized.Firstly,the K-means++ clustering algorithm is used to analyze the commodity dataset to obtain a suitable priori anchor to improve the location performance of the network.Then the cross-entropy loss function used by the confidence error and classification error is optimized by adding a weight factor,so that the network can get more sufficient and effective training by improving the attention to the inseparable samples.Finally,we enhance the ability to deal with occluded commodities of network by improving the non-maximum suppression algorithm.Through comparison experiments,it is proved that the improved YOLOv3 algorithm can effectively solve the problem of missed detection of the commodities with a large proportion and occluded commodities,and the overall detection performance has been improved to some extent.Due to the small sample size,light and other factors,the detection performance of small commodities has been reduced.In order to improve the positioning ability of the network and the detection performance of small commodities,according to the characteristics of the actual scene of supermarket shelf commodity detection,we design a better feature extraction network specifically for commodity detection,and propose an end-to-end commodity detection network CDNet(Commodity Detection Network).In order to fully extract the deep features of the image,the feature extraction network is designed with reference to the Residual Network,which uses multiple bottleneck residual blocks and 1 × 1 convolution to realize cross-channel information integration.In the second half of the feature extraction network,in order to retain a high-resolution feature map,the down-sampling is canceled to reduce the compression rate of image.In order to compensate for the lost receptive field of the high-resolution feature map,a dilated convolution is introduced to replace the bottleneck residual block with the dilated bottleneck residual block and the dilated residual block with 1 × 1 con projection.Finally,multi-scale prediction is introduced to fuse the features of different layers,so that the feature map contains different levels of resolution and semantic information,which can be used to detect different sizes of commodities.Through the comparative analysis of experimental results,the CDNet commodity detection network proposed in this paper can adapt to the detection of commodities from multiple perspectives,with an average recognition accuracy of 96.99% for all kinds of commodities and strong robustness,the positioning performance is greatly improved,and the missed rate of the small commodities is reduced.