强激光与粒子束2025，Vol.37Issue(12)：1-8,8.DOI:10.11884/HPLPB202537.250066
基于深度学习的小样本光学元件表面瑕疵识别
Few-shot defect recognition in optical components with deep learning
摘要
Abstract
[Background]Surface defects on optical components in high-power solid-state laser systems seriously impair the system's operational stability and laser output performance.However,precise detection of such defects under few-shot conditions remains a critical challenge,as limited training data often restricts the generalization ability of detection models and creates an urgent need for high-performance defect detection methods adapted to this scenario.[Purpose]To address this issue,this study aims to design and propose an enhanced detection method dubbed ICFNetV2,which is developed based on the existing ICFNet.Its core goal is to improve the accuracy and generalization of optical component surface defect detection under few-shot scenarios.[Methods]ICFNetV2 integrates data augmentation techniques with deep residual networks:Its framework adopts a synergistic design of residual connection mechanisms and decoupled channel convolution operations to construct a 34-layer cascaded network—this structure mitigates gradient decay during deep network training and enhances cross-layer feature transmission efficiency.The network also incorporates spatial dropout layers and implements a data preprocessing pipeline encompassing random rotation,mirror flipping,and Gaussian noise injection,which expands the training dataset to 9 times its original size.Additionally,ablation studies were conducted to verify the efficacy of each individual network module.[Results]Experimental results demonstrate that the optimized ICFNetV2 achieves a classification accuracy of 97.4%for three typical defect types,representing a 0.7%improvement over the baseline ICFNet model.[Conclusions]In conclusion,ICFNetV2 effectively enhances defect detection performance under few-shot conditions through architectural optimization and data augmentation.The validation from ablation studies and the observed accuracy gains confirm the effectiveness of its key modules,providing a reliable solution for surface defect detection of optical components in high-power solid-state laser systems and offering reference value for similar few-shot detection tasks.关键词
精密光学元件/深度学习/ResNet/缺陷分类/小样本/DropoutKey words
precision optical elements/deep learning/ResNet/imperfections classification/few-shot/Dropout分类
计算机与自动化引用本文复制引用
邵延华,忻晨,楚红雨..基于深度学习的小样本光学元件表面瑕疵识别[J].强激光与粒子束,2025,37(12):1-8,8.基金项目
国家自然科学基金项目(10976034、6160382) （10976034、6160382）
四川省科技厅项目(2019YJ0325) （2019YJ0325）
达州通川区科技项目(25YYJC0002) （25YYJC0002）
