Image feature extraction method, device, apparatus, and readable storage medium

Abstract

The invention discloses an image feature extraction method, which comprises the following steps: preprocessing an input image to be recognized to obtain an input feature map; setting the parameter ofa node in the convolution node on-chip network according to the node coefficients obtained in advance, splitting and packaging the input characteristic map data to obtain a communication data frame; sending the communication data frames to each node in the convolution node on-chip network in turn according to the corresponding node flow order for data processing to obtain output data; recombiningthe output data of each data frame to obtain the output data of the characteristic map, carrying out feature classification according to the output data of the feature map. The method divides the extracted feature map data into several data blocks and processes the data by routing, which can improve the efficiency of data transmission, reduce the operation time of convolution data and improve theefficiency of image recognition. The invention also discloses an image feature extraction device, an apparatus and a readable storage medium, which have the beneficial effects.

Description

technical field [0001] The invention relates to the field of convolutional neural network systems, in particular to an image feature extraction method, an image feature extraction device, an image feature extraction device, and a readable storage medium. Background technique [0002] In image recognition and classification, due to the superiority of convolutional neural network performance, convolutional neural networks are often used for image feature extraction. [0003] When hardware-accelerating the convolutional neural network, since each layer of neural network has multi-channel feature map input, filter coefficient input and corresponding multi-channel result output, in order to improve calculation speed and increase parallelism, the system Multiple computing cores are used for parallel computing. [0004] For convolutional neural networks, the input data is shared, and all output channels use these input data. At present, the input feature map data is generally pla...

AI Technical Summary

Problems solved by technology

[0005] When the above method is used for convolution data processing, a large number of computing cores will frequently queue up to access shared memory, forming bus and memory access conflicts, thereby affecting system efficiency

Images