Research And Application Of Cross-domain Data Joint Learning In Defect Detection

Nowadays,defect detection technology based on deep learning has been rapidly developed and applied,but it also faces some problems.First,there are few defect samples in the industrial field and the collection cost is too high,which makes defect detection face the problem of small samples.Network training is easy to overfit and lacks generalization ability.Second,the current defect detection networks only focuse on specific industrial scenarios for learning and training.For new application scenarios,rapid deployment cannot be carried out,in other words,the networks lack migration ability.Based on the above problems,we proposed the concept of cross-domain data joint learning,and tried to mine the value hidden in cross-domain data.This paper investigated and sorted out the public datasets in various industrial fields,analyzed the existing problems in the current defect detection datasets,and annotated the selected defect data,forming a new cross-domain defect detection dataset.It is used to verify the feasibility of cross-domain data joint learning.In this paper,a simple residuals cross-domain defect detection network is proposed,which implements the basic image dichotomy capability through simple residuals stacking.It is trained on the crossdomain defect detection dataset,which verifies the feasibility of crossdomain data joint training and shows good generalization ability.Then,a multi-scale residual cross-domain defect detection network was proposed based on ASPP module,and the learning ability of multi-scale features was added to adapt to the multi-scale characteristics of defects.Better defect detection performance is achieved.Finally,a two-stage cross-domain defect detection network is proposed based on segmentation network and classification network.The segmentation network is responsible for extracting defects from the complex texture of materials and locating them accurately.The classification network receives different levels of information from the segmentation network,uses the attention mechanism to focus on the defect itself,and realizes the binary classification task of whether the sample has defects.In addition,two detection methods,global detection and local detection,are used in the experiment,and their detection effects are analyzed.The two-stage cross-domain defect detection network is verified by cross-domain defect detection dataset and cross-domain defect detection confusion matrix.It shows excellent generalization ability and cross-domain migration ability,which proves the feasibility and wide application value of cross-domain data joint learning.