Image recognition method and device

Abstract

The invention discloses an image recognition method and device, and belongs to the technical field of machine learning. The method comprises the steps that in the image recognition process, when a layer of a deep neural network receives a feature map, acquiring at least one feature vector, located in the scanning window, of the feature map based on the scanning window corresponding to an intra-layer convolution kernel; and filtering the first target element in the at least one feature vector and the second target element in the at least one weight vector, the sum of the first target element and the second target element being 0, and the position of the first target element in the feature vector being the same as the position of the second target element in the weight vector; on the basis of the at least one filtered feature vector and the at least one filtered weight vector, carrying out convolution continuously; and based on the feature points obtained by convolution, obtaining a layer output feature map. According to the method, redundant operation in the convolution processing process is avoided, so that the calculation amount of image recognition is greatly reduced, and the image recognition speed and efficiency are improved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to an image recognition method and device. Background technique [0002] With the development of machine learning technology, in various application scenarios such as image search and product recommendation, computer equipment can recognize images based on deep neural networks. Specifically, the image to be recognized can be input into the deep neural network, and the deep neural network is operated layer by layer to output the recognition result of the image. [0003] The deep neural network includes multiple layers, each layer includes at least one convolution kernel (kernel), each convolution kernel is used to convolve the feature map (feature map) received by the layer, and output to the next layer feature map of . Among them, each convolution kernel can be regarded as a weight matrix of a certain size, each convolution kernel is composed of multiple weight elements, ...

AI Technical Summary

Problems solved by technology

[0007] The embodiment of the present invention provides an image recognition method and device, which can solve the problem of excessive calculation time caused by redundant operations in deep neural networks in related technologies

Images