Research And Application Of Cross-Domain Few-Shot Fine-Grained Image Analysis In Defect Object Detection

In recent years,with the continuous innovation of computer hardware and the continuous breakthrough of computing power,the research results of deep learning in computer vision have been successfully applied to various fields.In the field of industrial production,computer vision technology is often used for product defect detection.However,it is difficult to collect and label a large number of defect sample data.Furthermore,the characteristics of surface defects of various industrial products are too similar,which increases the difficulty of detection and identification.Therefore,it is of great practical significance to detect and identify a small number of defective industrial products with high precision.Combining with the characteristics of a small amount of sample data of industrial products with surface defects,this thesis first proposed a method based on data enhancement to expand the defect samples at the feature level and data volume perspective.Then,a defect recognition and target detection method based on cross-domain few shot learning and finegrained image analysis is proposed to solve the recognition and target detection on a small number of defect samples,which are personalized to the complex high-level semantic problems caused by cross-domain and the similarity of different surface defect features.Finally,experiments were carried out on the published datasets and the flange shaft defect dataset collected in this thesis,and the experiments demonstrate the effectiveness of the proposed method.The main research contents of this thesis are as follows:(1)In order to solve the problem of the small number of collected flange shaft surface defect samples,a data augmentation method was proposed,which based on supervised random combination.The method based on geometry,color,pixel points,etc.,which includes 18 supervised data enhancement strategies and randomly selects 4-9 methods for data enhancement of defective samples according to a certain probability.The experimental results of statistical distribution characteristics show that the random combinatorial data augmentation method expands the defect samples from the perspective of data volume and feature distribution.On the basis of data augmentation,the accuracy of image classification based on deep learning is improved from 66.7% to 93.3%.The results demonstrate the effectiveness of the method.(2)Although the defective samples can be expanded to a certain extent through data enhancement technology,it still cannot meet the training requirements of depth model.A cross-domain few shot defect recognition model for industrial parts is proposed,which is based on attention and adaptation.The residual attention mechanism is introduced in the feature encoder to solve the multi-dimensional feature distribution problem of cross-domain few shots.An adaptive bilinear matching network is built into the predictor to solve finegrained problems,which features such as shape,texture,or color are too similar between different defects.The experimental results show that the recognition accuracy of 5way-5shot on the public few shot dataset CUB is 63.78%;the recognition accuracy of 3way-5shot on the industrial defect dataset NEU-DET and flange shaft is 89.19% and 82.94%,respectively.The experimental results fully demonstrate the effectiveness of the method.(3)For the application scenario of defect target detection,a cross-domain small-sample defect target detection model,which based on multi-scale attention and dual Ro I Head is proposed.The model which based on the De FRCN model,integrates the feature pyramid network and attention mechanism into the backbone network to solve the multi-scale problem of defective samples and the problem that the cross-domain small-sample model can quickly learn during fine-tuning.Dual Ro I Heads were used to solve the fine-grained similarity problem of defect features in a region recognition network.The experimental results show that the n AP with strong comprehensive evaluation ability under 10-shot reaches 61.79% m AP.The model can achieve high-precision detection with a small number of defect samples.