Algorithm Research And System Implementation Of Electronic Shelf Label Detection

With the growth bottleneck arising in brick-and-mortar retailers and traditional ecommerce industry,the rise of “new retailing” has brought new energy into the sales market.As a real-time price display device,electronic shelf label becomes one of the most basic and important elements of “new retailing”.The automatic visual detection of electronic shelf label can locate goods and realize the automatic display of shelves,playing an important role in supporting the online and offline linkage of “new retailing” system.Existing object detection methods cannot detect the electronic shelf label very well.Although the deep learning object detection algorithm based on region proposal achieves high accuracy,it does not meet the real-time requirements.While,the algorithm based on regression method cannot extract the features of small and high-density objects well.Aiming at solving the fore mentioned problems,this thesis proposes a new feature fusion method,which can better fuse the location information and semantic information of the low-level feature map by fusing the adjacent feature map twice,and therefore improve the detection accuracy of high-density objects.Furthermore,the attention module is introduced to improve the detection effect of small object.The contribution of the thesis can be summarized as follow:(1)Production of massive electronic shelf label data set and design of detection system.This thesis competes the collection,filtering and annotation of the electronic shelf label data set,which contains 14713 pictures.In addition,according to the design requirements,this thesis designs an electronic shelf label detection system and describes all functional modules in detail.(2)According to the characteristics of electronic shelf label,such as small scale and high density,this thesis proposes a new feature fusion method.This method fuses features of adjacent feature map twice,which makes the network pay more attention to the location information and semantic information of the low-level feature map,so that the detection accuracy of small objects is improved.At the same time,the low-level feature map fused once overlaps the one fused twice,which makes the network pay more attention to highdensity objects and improves detection accuracy of high-density objects.Experiments prove that the performance of this method is significantly improved compared with SSD algorithm,and the m AP value is increased by 8.8%.(3)On the basis of feature fusion,this thesis adds attention mechanism to fused feature maps,as a result,the fusion process pays more attention to low-level feature map.Attention module further improves the detection accuracy of small objects.Experiments prove that the method,with attention module introduced,can prompt the performance of detection,and the m AP value is increased by 2.8%,totally the proposed method exceeds the baseline model by 11.6%.(4)The electronic shelf label detection system designed and implemented in this thesis is helpful to realize the online and offline linkage of the "new retailing" system.Besides,it completes the automatic location of goods and display of shelves.Therefore,electronic shelf label detection system has important economic and social value.29 figures,8 tables,53 reference articles.