Multi-label Image Recognition Method Based on Graph Attention Network

Abstract

The present invention is a multi-label image recognition method based on a graph attention network, comprising: the first step, the multi-label image to be recognized enters the convolutional neural network after being preprocessed by the input layer of the convolutional neural network, and utilizes the global co-occurrence The feature extraction module extracts the co-occurrence feature matrix; the second step is to use the conditional probability between the label nodes of the multi-label image to be recognized to construct the adjacency matrix between the label nodes; the third step is to use the adjacency matrix as a graph attention network The input of the graph attention network is used to obtain the learned matrix; the fourth step is to linearly multiply the co-occurrence feature matrix and the learned matrix to obtain the recognition result. This method uses the global co-occurrence feature extraction module to extract the co-occurrence feature matrix in the image, extracts the global co-occurrence feature and overall information of the image; calculates the correlation between the label nodes through the attention mechanism of the graph attention network, and calculates each label The nodes are adaptively assigned different weights, which is beneficial to improve the recognition accuracy.

Description

technical field [0001] The invention relates to the field of computer image processing, in particular to a multi-label image recognition method based on a graph attention network. Background technique [0002] Image recognition technology can replace manpower to process a large number of complex images. Image recognition is widely used in many fields, such as medical diagnosis, intelligent image management, and photo album search. [0003] In many image information processing, image recognition is actually a classification process, which is to find and identify the inherent features in the image to distinguish them from other images of different categories, which requires the selected features to be the most Distinctive features, the most discriminative features can be well distinguished from images of different categories, and this feature can describe the image vividly, that is, choose to have a small inter-class distance and a large class distance as much as possible The...

AI Technical Summary

Problems solved by technology

The graph structure model has more powerful representation capabilities and is more complex

Most of the existing graph-based multi-label image recognition models use pairwise compatibility probability or co-occurrence probability to establish the co-occurrence relationship between labels. However, these models usually cannot establish high-order relationships in images.

[0010] In short, the existing multi-label recognition methods based on deep learning neither fully consider the co-occurrence features between different objects, thus ignoring the correlation in categories, nor make full use of the high-order features in the image, reducing the multi-label recognition. Image Recognition Accuracy

Images