Supermarket commodity identification method based on deep learning

Abstract

The invention puts forward a supermarket commodity identification method based on deep learning, and solves the problem in the prior art that an identification rate is low in a real supermarket scene.The implementation scheme of the method comprises the following steps that: 1) manufacturing a supermarket goods shelf commodity training set; 2) constructing a commodity detection network, and carrying out training on the manufactured training set; 3) inputting a goods shelf picture into a trained network model to obtain all commodity target areas in a picture; 4) taking the output of a last convolutional layer in the network model as the characteristics of a commodity target; 5) coding the commodity characteristic to obtain a commodity descriptor; 6) calculating a similarity of the commodity model descriptor in an existing model library and the commodity descriptor; and 7) taking the most similar commodity model as an identification result. By use of the method, a commodity target areain the goods shelf picture can be accurately detected, in addition, the commodity target can be correctly identified, and the method can be used for managing supermarket goods shelf commodities.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a method for recognizing a shelf commodity target, which can be applied to supermarket shelf commodity management. Background technique [0002] In supermarkets, merchants and consumers need to obtain information about commodities on the shelves in real time. At present, the relevant information of these commodities is obtained manually, but the number of commodities in the supermarket is huge, and the way of manually obtaining commodity information is high in cost and low in efficiency. Therefore, the method of commodity recognition based on vision has important research significance and commercial value. [0003] The core of supermarket commodity recognition is target detection and recognition. Shaoqing Ren et al. published the paper "Faster R-CNN: Towards Real-Time Object Detection with Region ProposalNetworks" (IEEE Transactions on Pattern Analysis&Machin...

AI Technical Summary

Problems solved by technology

However, this method still cannot be applied to the identification of supermarket shelf products. The reasons are: 1. When this method performs ROI pooling on the feature vector of the candidate region, it will cause more target information loss. When the shelf product size is small, The recognition accuracy will be reduced; 2. This method is based on statistical recognition methods. There will often be new products in the supermarket and old products will be eliminated. Therefore, the classifier needs to be continuously retrained, which cannot meet the needs of real scenarios.

Images