RETAIL COMMODITY IDENTIFICATION BASED ON DIFFERENCE DETECTION

Aimed at the problems of poor robustness and low accuracy of difference detection design pattern in retail commodity target detection , an improved method is proposed. On the basis of DiffNet , the residual structure was used to fuse the difference features with the original features to enhance the ability of category information extraction. The twin network was combined with FPN structure to achieve the goal of multi-feature map prediction. Position attention mechanism was proposed to improve the use efficiency of target features. The detection strategy without anchor frame was used to reduce the model's dependence on the prior information of the data set , and the transfer learning method was combined to strengthen its ability to predict new categories. The experimental results show that the improved algorithm can better adapt to the complex detection environment , and the mAP and ACC reach the highest 98.1% and 91%.