High-resolution remote sensing image multi-label classification method of semantic multi-head attention mechanism

Abstract

The invention discloses a high-resolution remote sensing image multi-label classification method of a semantic multi-head attention mechanism. According to the method designed by the invention, the semantic region of the image can be positioned, and the relationship between the tags can be inferred autonomously. The network comprises three modules: a feature extraction module, a semantic sensitive module and a semantic relationship construction module. The feature extraction module extracts image features by using a deep convolutional network. The semantic sensitive module is used for locating semantic attention areas in the features and generating category expressions of content awareness. And the semantic relationship construction module is used for reasoning a label relationship in the content-perceived category expression to predict a final result. Experiments show that by means of the method, the semantic region can be effectively positioned, the relation between categories can be established with better robustness, and the method has higher precision in multi-label classification.

Description

technical field [0001] The invention belongs to the technical field of remote sensing image processing, is a semantic relationship learning method with a multi-head attention mechanism with local receptive fields, and can be applied to image multi-label classification. Background technique [0002] Due to the complexity of remote sensing scenarios, multi-label image classification of high-resolution remote sensing images is more general and practical than single-label image classification. High-resolution remote sensing scenes include various categories with correlations and differences among them. For example, for the correlation between categories, such as: "road" and "car" often co-occur in remote sensing images, "grass" and "water" accompany "golf course", and "airport" often contains "aircraft". There are not only semantic associations between categories, but also spatial location associations. Even if they have the same feature type, their spatial relationships are d...

AI Technical Summary

Problems solved by technology

Image feature extraction based on DCNN is an important step in image multi-label classification. However, due to the convolution operation of DCNN, the receptive field of DCNN is affected by the convolution kernel. The features extracted by DCNN mainly focus on short-distance relationships, and cannot focus on long-term distance relationship

However, the semantic learning method based on the multi-head attention mechanism can build a global relationship but cannot build a local receptive field.

Images