Image Classification Model and Its Classification Method Based on Hybrid Depth Separable Dilated Convolution

Abstract

The invention discloses an image classification model based on hybrid depth-separable dilation convolution, the construction process includes: from front to back, a depth-separable dilation convolution layer, a feature connection layer, a convolution layer, a batch normalization layer and a corrected linear unit The layer is packaged into a mixed depth separable dilated convolution module; from front to back, the convolution layer, batch normalization layer, corrected linear unit layer, mixed depth separable dilated convolution module, maximum pooling layer, flattening layer, The random deactivation layer and the fully connected layer encapsulate the backbone network of the deep neural network; the parameter weights of the backbone network are randomly initialized, and the number of iterations and the momentum parameters of the batch normalization layer are preset; the network model is optimized by the stochastic gradient descent method Parameters, iterative calculations are repeated until the loss value converges, and the optimal network model is obtained. Through the above solution, the present invention has the advantages of simple structure, less calculation workload, and accurate classification.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image classification model and a classification method based on hybrid depth separable dilated convolution. Background technique [0002] Images can objectively display natural things and are an important resource for people to understand the world, and technicians can obtain useful information and develop related algorithms by analyzing images. Image classification belongs to the direction of computer vision and is widely used in medical, food safety and other fields. [0003] At present, the main idea of ​​the image classification algorithm in the prior art is to assign corresponding labels to the image sets that need to be classified. For a computer, an image is a pixel matrix, which uses relevant algorithm technology to extract effective information in the pixel matrix. This is different from the way people perceive images. Traditional image classification algor...

AI Technical Summary

Problems solved by technology

[0008] First, this technology mainly fuses the features processed by depth separable convolution with a convolution kernel size of 3x3 and the features processed by ordinary convolution with a convolution kernel size of 1x1. Because the two parts have different feature receptive fields and scales Different, after complementation, the purpose of increasing the receptive field is achieved. Although the size of the receptive field is increased, the receptive field is still relatively single, and it is easy to ignore the classification of small targets;

[0009] Second, in this technology, the maximum pooling layer is used to downsample the feature map, but the resolution of the feature map is not repaired, the detailed information of the feature is greatly lost, and the classification accuracy is reduced;

[0010] Third, the parallel depth-wise separable convolution module mentioned in this technology improves the semantics of the feature map through stacked convolution, greatly increasing the parameter amount of the network structure, which is not easy to optimize, and does not combine context information for classification, which limits to a certain extent classification accuracy

Images