Fine-grained Commodity Identification For Large-scale And Dense Display Scene

With the rapid development of smart retail represented by unmanned supermarkets,electronic price tags,self-checkout,inspection robots,etc.,the demand for smart management such as display supervision,promotion control,distribution inspection,etc.,is also emerging gradually.Fine-grained identification and statistics of large-scale dense display goods can greatly speed up multiple production and operation processes,such as quantity statistics of inventory and display.However,few people focus on the research of the algorithm of commodity identification in large-scale intensive scenes.Therefore,we introduce a novel problem of fine-grained commodity identification in large-scale and dense display scenes.In this thesis,we make a detailed analysis of the optional challenges in this scenario and put forward a scenario-oriented commodity identification algorithm.Specifically,(1)In view of the goods in this scene have no background and the deformation is varied,we propose the Global Spatial Group module,which is effective in highlighting multiple active areas with various high-order semantics;(2)Considering a large number of categories of commodity data and the uneven degree of confusion among categories,the Focal Inter-Class Angular Loss is proposed to weight the similarity between-category pairs by the confusion matrix,so that the optimizer pays more attention to the difficult category pairs and can better extract the discriminant features;(3)Aiming at the characteristics of low resolution and large loss of visual information,the Super Resolution-Identification Joint Network was proposed,which enables the model to learn discriminative patterns for classification tasks from highdefinition images and recover its details from similar image blocks.In addition,this thesis constructs a commodity identification system in the aforementioned scenario and proves the superior performance of the proposed algorithm through extensive comparative experiments.