A multi-stream feature distance fusion system and a fusion method

Abstract

The invention discloses a multi-stream feature distance fusion system, which comprises a multi-stream feature extraction network, a contribution coefficient adaptive generation module and a distance fusion module which are connected step by step, wherein the multi-stream feature extraction network comprises a feature map extraction network used for extracting feature maps of input images and a feature extraction network used for extracting corresponding features from the feature maps respectively; the contribution coefficient adaptive generation module comprises an activation ratio calculationmodule and a contribution degree mapping module for calculating contribution coefficients of the regional characteristics according to the activation ratio of the regional characteristic patterns; and the distance fusion module is used for calculating feature distances between the corresponding feature maps and fusing the feature distances into multi-flow feature distances by utilizing contribution coefficients of the features. The invention also discloses a multi-stream feature distance fusion method. According to the method, the reference index which is more in line with the real conditioncan be provided for image similarity judgment through the multi-stream feature distance, and a breakthrough progress is brought to improving the accuracy of image similarity judgment.

Description

technical field

[0001] The invention relates to the field of image recognition, in particular to a multi-stream feature fusion system for image similarity judgment. Background technique

[0002] Image recognition usually requires image similarity calculation, and the feature distance of two images is generally used as the image similarity index. At present, most methods are dedicated to extracting a discriminative global feature and calculating the feature distance of the global feature. The smaller the feature distance, the higher the image similarity. However, in image recognition, it is often necessary to identify specific objects. Different objects may be similar on the whole, but different in local details. Therefore, regional features need to be further distinguished. The image recognition method based on global features adopted in the prior art cannot distinguish whether objects in images with a high overall similarity are the same target object. Contents of the in...

Images