Image classification method of deep learning network based on parallel asymmetric cavity convolution

Abstract

The invention discloses an image classification method of a deep learning network based on parallel asymmetric cavity convolution. According to the method, a parallel asymmetric cavity convolution module based on convolution parallelism is constructed by introducing asymmetric convolution. The module is composed of asymmetric convolution and dilated convolution, through the particularity of the asymmetric convolution structure, the module makes full use of information contained in a feature map under the condition that a receptive field is not changed, and the feature expression ability of a network model is improved. The parallel asymmetric cavity convolution module can be used for replacing traditional continuous convolution, and the accuracy of the whole model is improved under the condition that the complexity of the model is not increased. Any model embedded into the module is used for classifying the images, and the classification effect can be improved. The module embedding method is easy to implement, can be used for any model, and enables the model to have better robustness and accuracy.

Description

technical field [0001] The invention relates to the technical field of network pattern recognition, and relates to a parallel asymmetric atrous convolution module that can be embedded in any network structure, specifically an image classification method based on a parallel asymmetric atrous convolution deep learning network. Background technique [0002] Convolutional neural network is widely used in computer vision fields such as image classification, semantic segmentation and image generation. It uses the convolution kernel to convolve the local area in the image to extract the features in the image. In the convolution of each layer, pass The parameters are shared to reduce the complexity of the model, and then combined with the pooling operation to realize the identification of displacement invariance. Existing convolutional neural networks generally use 3×3 convolutions as their basic building blocks. For convolutional neural networks, the receptive field, depth, and nu...

AI Technical Summary

Problems solved by technology

[0003] The traditional convolution operation has three main disadvantages: first, it uses local operations, which cannot obtain a relatively large range or even the global features of the image, and the convolution kernel is a fixed size; second, it lacks adaptability to changes in the shape and posture of objects. Third, when the number of feature channels becomes larger, the parameters of the convolution kernel also become huge, increasing the amount of computation

At present, no one has proposed to use asymmetric convolution to solve the shortage of spatial level and information continuity of hole convolution.

Images