A convolutional neural network pooling method based on staggered rhombus perception

Abstract

The invention discloses a convolutional neural network pooling method based on staggered rhombus perception. The method comprises the following steps of adding blank rows and columns around a convolutional layer of a neural network; using a rhombic pooling window for replacing a square pooling window and slides from left to right from top to bottom, so that rows are overlapped and staggered; adopting a back propagation algorithm to train a convolutional neural network containing interlaced rhombus pooling, solving the gradient of each neuron, and obtaining a new weight. According to the present invention, the shape and the position of a local receptive field are changed; a pooling layer feature extraction mode is optimized, a norm is adopted to constrain a receptive field, a more effectivelocal receptive center is reserved, a staggered calculation mode is adopted, so that it is guaranteed that lines with the feature centers are partially overlapped, lines without the feature centers are exactly staggered, the asymmetry of the network is increased, and meanwhile the redundant calculation is reduced.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a convolutional neural network pooling method of interleaved diamond perception. Background technique [0002] Convolutional neural networks have the characteristics of local perception. Due to the invariance of translation, scaling and rotation of images, convolutional neural networks are widely used in image recognition, object detection and other fields. The correlation between image pixels is local, similar to how the human eye perceives small image blocks separately through the optic nerve, and makes a comprehensive judgment in the brain to obtain the characteristics of the entire image, without the need for each neuron to perceive the entire image. The local receptive field of the convolutional neural network is a feature extractor, and a local receptive field extracts a feature. Compared with the fully connected network, the convolutional neural network reduces t...

AI Technical Summary

Problems solved by technology

But most researchers limit their thinking to two

On the one hand, it is hoped to replace pooling with convolution with a step size greater than 1. As a result, 1*n plus n*1 convolution, hole convolution, deformable convolution, etc. have appeared to change the convolution method of the receptive field, but it has not changed. The rectangular shape of the receptive field, the effect is general

Images