Feature graph enhancement method and device of convolutional neural network, equipment and medium

Abstract

The invention discloses a feature map enhancement method and device of a convolutional neural network, equipment and a medium. The method comprises the following steps of performing convolution operation on an input original image; obtaining a corresponding multi-layer feature map; feature maps for a layer, grouping the data according to channel dimensions; obtaining multiple sub-feature maps, targeting each sub-feature map, performing global average pooling and global maximum pooling parallel processing by using an embedded spatial grouping enhanced SGE module; obtaining two corresponding channel dimension vectors; obtaining two corresponding channel dimension vectors according to the two corresponding channel dimension vectors; obtaining an attention enhancement factor of each channel inthe corresponding sub-feature map; according to the attention enhancement factor and the corresponding sub-feature map, obtaining the attention feature map obtaining the enhanced feature map corresponding to a certain layer of feature map according to all enhanced sub-feature maps, thereby better expressing the semantic information of the importance degree between sub-feature map channels, and improving the task performances of image classification, segmentation and detection of the convolutional neural network.

Description

technical field [0001] The present application relates to the technical field of computer vision, and in particular to a feature map enhancement method, device, equipment and medium of a convolutional neural network. Background technique [0002] With the rise of deep learning, as one of deep learning techniques, CNN (Convolutional Neural Network, Convolutional Neural Network) has been more and more developed and applied in the field of computer vision. Researchers have proposed many convolution operations, such as transpose Convolution, Dilated Convolution, Grouped Convolution, Depth-Separated Convolution, Pointwise Convolution, Deformable Convolution, etc. Among them, group convolution has great advantages in reducing the amount of calculations and parameters, preventing overfitting, etc., and is consistent with the grouping idea of ​​artificial design features in the early computer vision field, such as HOG (Histogram of OrientedGradient, histogram of oriented gradients) ...

AI Technical Summary

Problems solved by technology

[0004] However, the semantic feature information extracted by the existing SGE (Spatial Group-wise Enhance) module embedded in the CNN network structure is not fully expressed, which leads to the decline in the performance of CNN network image classification, segmentation, detection and other tasks

Images