Store comparison method based on image recognition

Abstract

The invention discloses a store comparison method based on image recognition. The store comparison method comprises a training stage: 1) forming m * k mini-batch store pictures; 2) the store picture passing through the last convolution layer of the neural network to obtain a feature map with the size of (N, C, H, W); performing global pooling to obtain global features of the store pictures, usinga tri-hard strategy to construct a triple, and using a triple loss function to carry out training; 3) obtaining a concerned local region, then calculating a corresponding region on the feature map, obtaining local features of the store picture through global pooling, and training by using the local features and the triple loss function; an inference stage: combining the global features and the local features of the store pictures to obtain final features expressed by the store pictures; and calculating the distance between the final features of the two store pictures, if the distance is less than a specified threshold, determining that the two store pictures are the same store picture, otherwise, determining that the two store pictures are not the same store picture. The method has the advantage of improving the recognition effect of the store.

Description

technical field [0001] The present invention relates to the technical field of store identification, in particular to a technique of store comparison method based on image recognition. Background technique [0002] Most of the existing technologies are based on deep learning, using the neural network to scale the original image or the original image to different sizes, using the softmax loss function to learn a global feature, and using this feature for subsequent retrieval or comparison Task. However, in the actual scene, the angle and light of the store photos taken by users change greatly, which increases the difficulty of identification. Some photos even contain a lot of non-store background information, such as the road, other nearby stores, the sky, etc. It is difficult to get a good description of the store by directly using global features, which will affect the final recognition result. Contents of the invention [0003] In order to overcome the deficiencies of ...

AI Technical Summary

Problems solved by technology

However, in the actual scene, the angle and light of the store photos taken by the user change greatly, which increases the difficulty of identification

Some photos even contain a lot of non-store background information, such as roads, other nearby stores, the sky, etc. It is difficult to get a good description of the store directly using global features, which affects the final recognition result

Images