Multi-scale fine-grained image recognition method and system based on multi-granularity attention

Abstract

The invention belongs to the technical field of image processing, and discloses a multi-scale fine-grained image recognition method and system based on multi-granularity attention. The method comprises the following steps: constructing an attention-based multi-granularity structure, dividing a feature extraction network into a plurality of stages, inputting images with different granularities into different stages of the feature extraction network, and carrying out feature extraction on the image to obtain an original feature map; obtaining attention weights from a channel domain and a space domain for the feature map of each stage through a multi-granularity attention module, fusing the attention weights, and then carrying out weighted fusion on the attention weights and the feature map to obtain key regions of different granularities in different stages; constructing a parallel multi-scale convolution module, grouping the feature maps, independently applying different types of convolution kernels to each group of feature maps, and performing feature extraction on the feature maps with different scales and granularities in different stages; and finally, carrying out feature fusion on the obtained feature map. The relationship between different regions can be fully mined, and low-dimensional space information and high-dimensional semantic information are fused.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to deep learning and fine-grained image recognition technology, in particular to a multi-scale fine-grained image recognition method and system based on multi-granularity attention. Background technique [0002] Fine-grained image recognition aims to classify more detailed subclasses of coarse-grained categories. However, fine-grained image recognition remains a challenging task due to the high intra-class variance and low between-class variance of fine-grained images. [0003] Early fine-grained image recognition methods addressed this issue with part-based feature representation via human-annotated bounding boxes / part annotations. However, specialized knowledge and a lot of annotation time are required in the labeling process. Therefore, for practical fine-grained image recognition tasks, strongly supervised methods that require a lot of time and resources for annotation ...

AI Technical Summary

Problems solved by technology

[0005] In view of the deficiencies in the above-mentioned existing technologies, the present invention provides a multi-scale fine-grained image recognition method and system based on multi-granularity attention. (1) First, in order to solve the problem of how to locate key areas in different image granularity by the attention mechanism, An attention-based multi-granularity structure is proposed, which divides the feature extraction network into several stages. Through the multi-granularity attention module, each stage focuses on capturing the most discriminative features with specific granularity in the corresponding layer of the feature extraction network. area; (2) Then it is proposed to use the parallel multi-scale convolution module to extract feature maps of different scales and different granularities in different stages. The module is divided into several levels, each level contains convolution kernels of different sizes and depths, parallel multi-scale The convolution module can use convolution kernels of different sizes to process input feature maps in parallel to capture the details of feature maps of different scales and granularity; (3) Finally, for the problem of fusing features from different regions in a synergistic manner, this paper Invented and designed a feature fusion module, the feature maps of different stages are compressed into feature vectors through the maximum pooling layer and fused, which not only fully excavates the relationship between different regions, but also effectively integrates low-dimensional spatial information and high-dimensional semantic information , thus improving the classification accuracy

Images