A Feature Map Processing Method Based on Hierarchical Group Convolutional Neural Network

Abstract

A method for processing feature maps based on hierarchical group convolutional neural networks, comprising: grouping input feature maps along the channel dimension to obtain several groups of first-generation sub-feature maps; The first 1×1 convolution processing is performed to obtain the first group of second-generation sub-feature maps; the channel selection operation is performed on the first group of second-generation sub-feature maps to obtain the first group of third-generation sub-feature maps; sequentially from the second group As the last group, splicing and convolution processing is performed on the first-generation sub-feature maps of each group; the second-generation sub-feature maps of each group are stitched together along the channel dimension to obtain the output feature map. In the method of the present invention, the input feature maps are grouped and processed, and the channel selection mask obtained after the channel selection process is used to represent whether the channel is selected as the feature map that needs to be spliced ​​for subsequent grouping, which reduces the number of features of the previous group that are fused into subsequent groupings. Graph channel dependencies generate redundant information, making convolution computation more efficient.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a feature map processing method based on a hierarchical group convolutional neural network. Background technique [0002] In recent years, convolutional neural network (CNN) has made significant progress in many fields, benefiting from excellent performance, mainly because CNN can learn complex nonlinear mapping relationship between input and output through a large amount of training data. In the CNN stacked convolution module, all channels generated by the previous layer are treated equally by the next layer. This uniform distribution may not be optimal, because for a certain layer, some features may be more important than others. This is especially true for the upper layers of CNN, where only a small number of channels are activated, and the activation values ​​of neurons in other channels are close to zero. [0003] At present, many lightweight networks use grouped co...

AI Technical Summary

Problems solved by technology

[0005] like figure 1 As shown, in the prior art, 1×1 convolution is used for Hierarchical Group Convolution (HGC) to process each group of feature maps and then directly spliced ​​to the next group as input feature maps. This operation is redundant. The remaining information ignores the distribution law that only a few of the high-level channels are activated, which reduces the computational efficiency of the convolution module in the HGC structure to a certain extent.

Images