Shelf Commodity Detection Based On Deep Learning

In the retail industry,in order to manage the distribution channels better,enterprises need to regularly inspect the display information of goods on shelves in offline stores.however,the traditional manual inspection method is costly.Using the commodity detection algorithm can automatically analyze the categories and quantities of commodities in the images of shlevs instead of manpower.It can greatly improve the efficiency of the whole retail industry.In different store scenes,such problems as the change of commodity size,the tilt of shooting angle and the change of illumination intensity will lead to the complexity of image background.Traditional commodity target detection algorithms can not cope with these complex changes.This paper studies these problems.In this paper,the main research work and results are summarized as follows:(1)The dataset of special commodity image has been established.The dataset of commodities taken by supermarkets and small stores all over the country.All the photos were taken under different light intensity,different shooting distance and different shooting angles.(2)Research on object detection algorithm.The most popular framework of Faster R-CNN for target detection is studied.The detection effect of Faster R-CNN in the application of commodity detection task and the factors affecting the detection performance are analyzed.(3)This paper solves the existing problems in two ways according to the particularity of the practical application scenes: data augmentation and network structure improvement.The basic idea of data augmentation is to make the size and type distribution of data more balanced by generating false data.The basic idea of network structure improvement is to enhance the ability of the network to express small targets and deformed targets caused by inclined shooting.Combined with the network structure of cascaded network,characteristic pyramid and deformable convolution,the object detection model based on Faster-RCNN has been improved,and the accuracy of commodity detection reaches 0.92 mAP.