Device and method for multi-scale feature extraction of weld defects

Abstract

The invention discloses a multi-scale feature extraction module and method for weld defects based on lightweight hollow convolution, which belongs to the technical field of weld defect detection. The road includes connected lightweight dilated convolutions and spatial pooling pyramids. Lightweight dilated convolutions are used to extract weld defect features of different receptive fields from the weld defect feature map to generate the first feature map. The spatial pooling pyramid is used to The second feature map is generated after enhancing the expressive ability of the first feature map for weld defect features of different scales through pooling layers of different sizes; the expansion rate of the lightweight atrous convolution of different branches is different; the first superposition layer uses The second feature map output by all branches is superimposed to generate the final feature map. The invention improves the weight of the model, enhances the robustness of the neural network model against large changes in the size of weld defects, and can be conveniently and quickly embedded into various models.

Description

technical field [0001] The invention belongs to the related technical field of weld defect detection, and more particularly, relates to a multi-scale feature extraction module and method for weld defects based on lightweight hollow convolution. Background technique [0002] There are many types of weld defects, and the shape and size are also very different. It is difficult to learn the feature information that can deal with different defect types at the same time when extracting them in the convolutional neural network model. In the case of large differences in target size, the information extracted by the neural network model needs to have different receptive fields and multi-scale feature sources. In addition, in industrial applications, the complexity of the model needs to be considered, and the model should be lightweighted as much as possible to improve production efficiency. [0003] The patent (application number 202011206236.2) discloses a method for detecting smal...

AI Technical Summary

Problems solved by technology

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a multi-scale feature extraction module and method for weld defects based on lightweight hollow convolution, thereby solving how to make the weld defect information extracted by the module have different Receptive fields and multi-scale feature sources, while making lightweight improvements to the technical issues of the module

Images