Efficient coding method and device for convolutional neural network, and storage medium

Abstract

An efficient coding method of a convolutional neural network comprises the following steps: calculating a first frame of image by using the convolutional neural network, and selecting and storing specific row and column data of a middle layer as reference convolution; when a first layer of convolution of each subsequent frame of image is calculated, comparing the first layer of convolution with a reference convolution, and outputting convolution results of a plurality of channels with small changes or unchanged to carry out run-length coding; when data are input according to the run-length coding area and the non-run-length coding area, specific row and column data of the corresponding middle layer are taken out to execute convolution operation of different strategies; and updating the reference convolution every preset frame number of images. According to the method, the requirement of importing and exporting intermediate layer data between the NPU and the memory on transmission bandwidth is greatly reduced, repeated calculation of the convolutional neural network on a static image area can be effectively reduced, the calculation speed of the neural network is improved, the requirement of real-time response is met, and meanwhile, the power consumption is greatly reduced.

Description

technical field [0001] The invention relates to the technical field of processing images by using a convolutional neural network, in particular to a high-efficiency coding method, device, and storage medium for a convolutional neural network. Background technique [0002] Convolutional neural networks have been widely used in image recognition. Due to the large amount of calculation, there are great challenges to the area, price and power consumption of the chip. In real scenes, the background of the video captured by the camera is mostly static. If the convolutional neural network can reduce the repeated calculation of the slightly changed image area of ​​​​multi-frame images, the calculation speed of the convolutional neural network will be greatly improved and the calculation speed will be reduced. power consumption, and meet the needs of many real-time application scenarios. [0003] In the prior art, because parts of the first frame image and the subsequent frame image...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, due to the fact that the first frame image and the subsequent frame image are similar or identical, repeated calculations lead to a huge amount of data transmission in the middle layer of the convolution, and at the same time reduce the speed of the convolution operation.

Images