Image processing method based on winograd dynamic convolution block

Abstract

The invention discloses an image processing method based on winograd dynamic convolution blocks, and belongs to the field of convolution networks. According to the invention, a calculation complexityfunction of a winograd rapid convolution method is generated by using a Chinese remainder theorem algorithm, and convolution parameters of each layer in a convolutional neural network model are introduced as constants by the calculation complexity function to obtain a calculation complexity model of which the variable is the size of a winograd convolution block; the method also includes minimizingthe computation overhead based on the computation complexity model; according to the convolution block size obtained by minimizing the calculation overhead, completing winograd rapid convolution calculation of the corresponding layer number; extracting features of the pictures and sending the features to a convolutional neural network for classification processing; according to the invention, theproblem of convolution performance reduction caused by unbalanced efficiency of a single convolution block of a winograd algorithm on a general computing platform is solved, and the image processingmethod can accelerate the computation of the convolutional neural network computed by the processor.

Description

technical field [0001] The invention belongs to the field of convolution network, in particular to an image processing method based on winograd dynamic convolution block. Background technique [0002] Convolutional neural network (CNN) is a set of deep learning algorithms that perform well on a variety of AI tasks, including video surveillance, speech recognition, natural language processing, and autonomous driving. Over the past decade, CNNs have shown great promise and been the focus of a great deal of research. Convolutional layers are memory-intensive, computationally intensive, and prevalent in many advanced CNNs, including AlexNet, VGG, OverFeat, and ResNet. Therefore, the convolutional layer is the main factor affecting the overall performance of CNN. [0003] Massive datasets and more complex models can provide satisfactory results and significantly improve the final accuracy of the task. However, it also results in increased training overhead and more computation...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to solve the problem of the convolution performance decline caused by the unbalanced efficiency of a single c...

Images