Channel extension method based on convolutional neural network

Abstract

The invention provides a channel extension method based on a convolutional neural network. The method is used for replacing a general convolution extension channel and pooling operation in the convolutional neural network, so the redundant data of intermediate operations of the convolutional neural network can be reduced without losing valid data. The method is characterized by comprising the following steps: step S1, calculating the resolution and the channel number of feature map data according to the resolution change rate of feature maps before and after convolution pooling, and calculating the channel number of the feature map data according to the channel number of the feature map data after convolution pooling; s2, performing dimension splitting on channels with corresponding heights and widths in the feature map data of the four dimensions to form six dimensions; s3, carrying out dimension turning on the split feature map data, wherein the third dimension and the fifth dimension are changed into the fifth dimension and the sixth dimension; and S4, carrying out dimension combination on the second dimension, the third dimension and the fourth dimension in the changed featuremap data, and re-integrating into four dimensions.

Description

technical field [0001] The invention belongs to the field of deep learning and computer vision research, and in particular relates to a channel expansion method based on a convolutional neural network. Background technique [0002] Digital image analysis technology plays an important role in today's society. The development of machine vision technology has gradually abandoned the manual design algorithm of traditional digital image processing, and turned to deep learning, with Convolutional Neural Network (CNN) as the Representative, in order to achieve high accuracy target detection results. However, the resolution of the image data directly processed by the existing CNN network is low, which is far from compatible with the commonly used high-definition images. One of the reasons is that when the feature extraction algorithm is executed in the CNN model, the feature maps (feature maps) The number of channels is huge, and there is too much redundant data in the middle. The...

AI Technical Summary

Problems solved by technology

However, the redundant data brought about by channel expansion in this process will inevitably consume huge computing resources. When the image data is large or the depth model is complex, the computing resources of memory / video memory will not be able to bear the redundant data in the computing process. surplus data, this huge resource consumption can become an intractable problem

Images