Convolution neural network algorithm optimization method and device based on Neon instruction

Abstract

The invention provides a convolution neural network algorithm optimization method based on the Neon instruction. The method comprises the steps that matrix processing is carried out on the convolutionkernel image of a convolution layer to acquire a corresponding A matrix, and the A matrix columns are aligned according to the multiple of four; a to-be-convoluted image is input; matrix processing is carried out on the input image to be convoluted to acquire a corresponding B matrix; the B matrix rows are aligned according to the multiple of four; the B matrix is transposed to acquire a transposed matrix Bt; the row and row dot product of the A matrix and the Bt matrix is calculated; and the Neon instruction is used for parallel optimization. Compared with the prior art, the method providedby the invention can effectively improve the computing performance of the convolution neural network.

Description

technical field [0001] The invention relates to image processing, video monitoring and convolutional neural networks, in particular to an algorithm optimization method and device for convolutional neural networks based on Neon instructions. Background technique [0002] With the rapid development of artificial intelligence, deep learning has been increasingly introduced into the fields of image processing and pattern recognition, and has performed well in solving related problems. Among them, convolutional neural networks (CNN for short), as a model structure of deep learning, are especially good at dealing with machine learning problems related to images, especially large images, and have been widely used and the most in-depth research. [0003] However, in the practical application of image processing and pattern recognition, the convolutional neural network is generally implemented with more network layers, and its computational complexity is relatively high, including a ...

AI Technical Summary

Problems solved by technology

However, for some embedded ARM platforms that do not support third-party matrix operation acceleration libraries, the convolutional neural network algorithm still takes a long time and the real-time performance is not good.

Images