Research On Surface Defect Recognition For Optical Components Based On Lightweight Deep Learning Architecture

In the optical industry,accurate identification of surface defects is critical for the rapid screening of defective parts.However,traditional surface defect recognition algorithms rely heavily on manually constructed defect features,which can be limited in their ability to differentiate weak and complex defects due to the inherent complexity and weakness of optical part surface defects.In recent years,deep learning technology has emerged as a promising tool for defect detection due to its ability to automatically extract features and its end-to-end architecture.Nevertheless,existing deep learning-based optical surface defect recognition algorithms still face challenges,including limited coverage of defect types and difficulty in achieving a good balance between accuracy and speed.To address these challenges,this thesis proposes a lightweight optical part surface defect classification model based on deep neural networks.The proposed model achieves accurate identification of four types of manufacturing defects and two types of non-manufacturing defects in optical parts.The main research contents are as follows:(1)Establishing a Database of Surface Defects for Optical Components:The surface defect data of optical components were collected and preprocessed,resulting in a total of 8500 defect images.For subsequent experiments,these images were divided into training,validation,and testing sets in a ratio of 3:1:1,serving as the training,validation,and testing data for the optical component surface defect recognition model.This work was conducted at a doctoral level.(2)Proposing a Deep Learning-Based Classification Method for Surface Defects in Optical Components:To enhance the capability of convolutional neural networks in focusing on crucial defect information,a multi-scale mixed convolutional kernel and an asymmetric mixed convolutional kernel were designed.Based on these two mixed kernels,the M~2-Net model was constructed for the classification of surface defects in optical components.This model achieved fine discrimination among six categories,including manufacturing defects,non-manufacturing defects,and defect-free cases.This research was conducted at a doctoral level.(3)To further enhance the practical inference efficiency of the model,this study introduces the concept of structural reparameterization and proposes the lightweight M~2Rep-Net model.By applying structural reparameterization to the branches of M~2-Net’s multi-scale mixed convolutional kernel and asymmetric mixed convolutional kernel,the model is transformed into an equivalent single-branch structure.This directly accelerates the actual inference speed while maintaining the same recognition performance.This research was conducted at a doctoral level.(4)To further improve the recognition efficiency of the model,this study introduces the L1-norm pruning strategy to optimize the M~2Rep-Net model.By removing redundant convolutional filters with lower importance for optical surface defect detection,the model is compressed in size and accelerated in computation.Simultaneously,the model’s accuracy loss is kept within an acceptable range.This research was conducted at a doctoral level.