Feature extraction method, and image processing method and device

Abstract

The invention provides a feature extraction method, and an image processing method and device. The feature extraction method comprises the steps: obtaining a feature map, performing deformation and convolution on the tensor of the feature map through a first size adjusting step, a first convolution step and a second size adjusting step, performing further deformation and convolution on the tensorof the feature map through a third size adjusting step, a second convolution step and a fourth size adjusting step, and finally obtaining features of the feature map through a feature extraction step.According to the feature extraction method provided by the invention, the tensor is deformed and combined in the convolutional neural network convolutional layer, so that the convolution calculationamount is reduced, and the operation efficiency is improved.

Description

technical field [0001] The present invention generally relates to the field of image recognition, and specifically relates to a feature extraction method, an image processing method and a device. Background technique [0002] With the development of computer technology, more and more scenes need to use computer technology to perform image processing tasks such as target detection and target recognition. Among them, the convolutional neural network (CNN) model is the core of modern deep visual recognition systems. However, since convolutional networks usually contain a huge amount of computation, they must be simplified in order to apply the model to low-power scenarios such as mobile devices. Among the many simplification strategies, depthwise separable convolution is one of the commonly used techniques. It decomposes the convolutional layer into a combination of depthwise convolution and pointwise convolution. Reduce the computational load of the model. [0003] The curr...

AI Technical Summary

Problems solved by technology

[0003] The current efficient model structure still has the following problems: 1) The theoretical calculation amount and parameter amount of depth convolution are very low, but the overall efficiency is not high due to its low calculation density in actual operation; 2) Group convolution ( groupconvolution) as a method to reduce the amount of calculation parameters, also due to the characteristics of low calculation density and high calculation fragmentation, it is often difficult to achieve ideal efficiency in practice; 3) In other structures, there are some difficult-to-optimize operations , such as tensor dimension shuffle / transpose, has become a bottleneck that prevents algorithms from being deployed on hardware to achieve fast operations

Images