Pyramid pooling multi-scale feature learning method adopting attention mechanism

Abstract

The invention discloses a pyramid pooling multi-scale feature learning method adopting an attention mechanism. The pyramid pooling multi-scale feature learning method is characterized by comprising apyramid pooling model and an attention model. The pyramid pooling model is formed by connecting a plurality of self-adaptive pooling paths with different pooling window sizes in parallel, and is usedfor extracting multi-scale context features. The attention model represents the relationship between different channel features generated by the pyramid pooling model through a nonlinear function, andallocates weights to the channels to reinforce useful features and suppress useless features. According to the pyramid pooling multi-scale feature learning method adopting the attention mechanism, the multi-scale feature learning capability of a pyramid pooling model can be enhanced, the method can be flexibly embedded into a convolutional neural network model, and the method is suitable for various tasks such as image semantic segmentation, target detection and image classification.

Description

technical field [0001] The invention belongs to the field of deep learning, in particular to a pyramid pooling multi-scale feature learning method using an attention mechanism. Background technique [0002] Multi-scale features are often used in deep convolutional networks to improve network performance. As the most commonly used multi-scale feature extraction model, the pyramid pooling model extracts multi-scale context features by paralleling multiple adaptive pooling channels with different pooling window sizes, and then uses 1×1 convolution to linearly fuse the features of each channel . However, the use of linear functions alone is not enough to describe the nonlinear relationship between channels of multi-scale features, resulting in the inability of the pyramid pooling model to effectively extract multi-scale features. Contents of the invention [0003] In order to overcome the problem of insufficient nonlinear representation ability of the above-mentioned pyramid...

AI Technical Summary

Problems solved by technology

[0003] In order to overcome the problem of insufficient nonlinear representation ability of the above-mentioned pyramid pooling model, the present invention provides a pyramid pooling multi-scale feature learning method using an attention mechanism

Images