Research On Deep Learning-Based Surface Defect Detection Methods

Industrial surface defect detection refers to the use of various techniques and equipment to detect defects on the surfaces of industrial products in order to ensure product quality and production efficiency.In complex industrial surface defect detection environments,manual inspection methods are difficult to cope with and can be costly.With the development of machine vision technology,a large number of machine learning and deep learning methods have been applied to industrial surface defect detection.These methods offer higher accuracy,efficiency,and lower costs compared to manual inspection methods.However,machine learning methods typically require a large number of data samples,and a lack of sufficient data can lead to overfitting of the training model and difficulties in achieving good generalization.Additionally,in practical detection scenarios,imbalanced defect data categories are often encountered,which can cause the trained models to be biased towards common categories and result in distorted evaluation metrics.To address these issues,the paper proposes three defect detection models and validates their detection performance on six surface defect datasets.The specific work is as follows:(1)Propose a transfer learning-based surface defect classification method.This method first pretrains a deep convolutional neural network VGG16 on the Image Net dataset and then adjusts and fine-tunes the network structure for the target dataset.The classification performance of the model is tested on a dataset of magnetic tiles,demonstrating its effectiveness.The accuracy of recognizing six classes of magnetic tile defects reaches 98.69%,significantly higher than manual classification accuracy and traditional machine vision classification methods.The experimental results show that this method achieves high accuracy in magnetic tile defect classification while greatly reducing training time.(2)Design a surface defect detection framework based on generative adversarial networks(GANs),using GANs to generate synthetic images for data augmentation.Two GAN architectures,namely Deep Convolutional Generative Adversarial Networks(DCGAN)and Auxiliary Classifier Generative Adversarial Networks(ACGAN)with an auxiliary classifier,are trained and compared with traditional data augmentation methods.The entire data augmentation process is implemented on the NEU-CLS dataset of surface defects on hot-rolled steel from Northeastern University.The recall rate and precision rate of hot-rolled steel surface defect classification reach 95.33% and 99.16%,respectively.The experimental results demonstrate that this method significantly improves the performance of convolutional neural networks in surface defect classification and produces models with strong generalization capability.(3)Design a pixel-level segmentation and image-level classification network.The model consists of three-stage network architectures,which can extract key features,spatial position information,and semantic information,and perform defect segmentation and image classification tasks.Adopting a two-stage training mode ensures that the parameters of the segmentation network and classification network are unconstrained and do not lead to confusion or non-convergence.The loss function is improved to enable fast and accurate convergence of parameters.The model is validated on three surface defect datasets,achieving good detection results.The network can quickly and accurately classify and segment defects.The paper primarily focuses on machine vision-based industrial surface defect detection methods,providing a reliable basis for practical applications in industrial production.It holds significant theoretical and practical value in the field of machine vision-based industrial product manufacturing technology.