Fine-grained image recognition method and system based on block detector and feature fusion convolutional neural network

Abstract

The invention discloses a fine-grained image recognition method and system based on block detector and feature fusion convolutional neural network. The method comprises the following steps of: firstly, obtaining local features and global features; directly classifying the global features to obtain loss1; respectively filtering the local features and the global features by using a block detector toobtain local filtering features and global filtering features; carrying out global maximum pooling on the obtained local filtering features to obtain local discriminative feature blocks, wherein thelocal discriminative feature blocks can be directly classified to obtain loss2; in addition, constructing a feature fusion flow to fuse the local filtering features and the global filtering features,obtaining a hierarchical multi-layer representation, and obtaining loss3 through direct classification; introducing an attention cross-layer pooling method to carry out filtering supervision on the network to obtain a loss4; and finally, performing weighted summation on the four loses to obtain total loses, so that fine-grained image recognition can be effectively realized.

Description

technical field [0001] The invention belongs to the technical field of computer vision and pattern recognition, in particular to a fine-grained image recognition method and system based on a convolutional neural network based on a block detector and feature fusion. Background technique [0002] With the ever-increasing image data, image processing technology is becoming more and more important in modern life. Among them, the fine-grained image recognition problem has more and more theoretical research value and practical application value. [0003] At present, the existing fine-grained recognition problems mainly have the following problems: (1) Some methods use local localization networks to learn discriminative local feature blocks, but such methods rely on a large amount of additional artificial part labeling information, which consumes a lot of time. A lot of time and resources, and manual labeling does not have good objectivity; (2) Another part of the method embeds di...

AI Technical Summary

Problems solved by technology

[0003] At present, the existing fine-grained recognition problems mainly have the following problems: (1) Some methods use local localization networks to learn discriminative local feature blocks, but such methods rely on a large amount of additional artificial part labeling information, which consumes a lot of time. A lot of time and resources, and manual labeling does not have good objectivity; (2) Another part of the method embeds different sub-networks in the main network structure to obtain different fine-grained feature representations, thereby assisting the main network to learn more Good fine-grained features. Although these methods have achieved good results, they need to optimize the sub-network in turn, and the optimization and adjustment of the network is more complicated; (3) the current convolutional neural network in fine-grained recognition tasks almost only uses low-level features. This is not enough feature representation for fine-grained image classification, so there is still a lot of room for improvement in mining more advanced features

Images