Research On Surface Micro-defect Detection Technology Based On Wavelet Transform And Deep Learning

Currently,the detection of surface micro-defects is a prominent issue in industrial research.This thesis focuses on the study of defect detection based on deep learning,with a specific emphasis on the detection of surface micro-defects.Considering the challenges associated with deep learning-based surface micro-defect detection,such as high cost,high false positive rates,and low accuracy,this thesis proposes improvements to existing neural network architectures and presents two innovative methods for micro-defect detection.The main contributions of this thesis are as follows:Firstly,this thesis introduces a novel deep learning network structure,infused with a wavelet transform module.This architecture integrates a "Wavelet Convolutional Fusion Layer" and a "Wavelet Transform Layer",incorporating sampling and convolution operations preceding the wavelet transformation operations.The intent behind this is to augment the network’s perceptibility of the target and to optimise detection outcomes.When applied to the detection of scratches and dot defects in surface defect datasets,the network with the new structure achieves an impressive m AP(mean average precision)of over 90%.Furthermore,to evaluate the network’s detection performance in complex environments,we conduct target detection experiments using aerospace datasets.The results show a remarkable improvement of nearly 15% in the detection rate of small targets compared to the original network.These findings provide strong evidence that this network structure significantly enhances the detection accuracy of micro-defects.Secondly,this thesis advances a feature scale transformation technique premised on background backtracking.This technique manipulates the prediction frame,ensuring the feature information is as much skewed towards the possible scale of the target as possible,thus enabling swifter detection.This method employs confidence scoring to determine whether the feature point constitutes the background,and specific operational strategies have been engineered to reduce the influence of background features on higher-layer features.To alleviate the computational load,this thesis also designs a vector zero assignment strategy.This approach is employed within this thesis to experiment with targets of varying scales within a surface defect dataset.The refined network yielded an approximately 20% increase in m AP for point defect detection and an approximate 11%improvement in m AP across targets of varying scales.These experimental results indicate that compared to the original network,the new techniques proposed within this thesis elevate the detection accuracy of varying scale targets,without compromising detection speed.Lastly,in response to the scenario of mobile phone screen defect detection,this thesis has developed a practical detection software system that encompasses data management,defect detection,and model management functionalities,thereby enabling the application of these research outcomes in real-world scenarios.Overall,this thesis introduces two innovative solutions to the issue of small target detection,providing fresh perspectives and potential paths for improvement for related small target detection endeavours.