Research On Fabric Defect Detection Technology Based On Deep Learning

Defect detection in fabric is a necessary step in quality control for industries related to fabrics such as clothing and tents.Traditional fabric inspection relies on human eye methods,which are inefficient and inaccurate.In this article,based on the actual situation,we designed an unsupervised model to detect VIT-AD,and established two main optimization research directions for data enhancement and network models.The former aims to reduce the false positive rate of fabric images affected by lighting,folds and other environmental factors,while the latter aims to map feature distributions reasonably onto normal distributions,thus exposing the distribution of defect features and achieving effective defect detection and localization.The main work is as follows:Firstly,in terms of data sets,in order to train high-performance network models,this paper collects and organizes real fabric defect data sets for the special use scenarios of factories.In order to verify the generalization performance of the algorithm,this paper also organizes public general defect benchmarks as the main test objects.Secondly,in terms of data augmentation,this paper proposes the Cut&Move approach,which involves cutting out specific regions of an image and moving them to different positions to simulate irregularities such as lighting and wrinkles.This effectively alleviates the problem of normal samples being falsely detected as defective samples due to environmental factors.Experimental results demonstrate a 3.0%improvement in AUROC on an industrial fabric defect dataset.Thirdly,in terms of network model,this paper proposes the VIT-AD algorithm.To obtain more global and detailed local information features,a CNN cascade structure or a Transformer structure with self-attention is used,along with a2D-Flow structure and double loss function.This ensures that the feature distribution extracted from the defect area during the detection phase is more stable and reliable.Through experiments,VIT-AD achieved a95.0%detection rate on an industrial fabric defect dataset and demonstrated excellent detection performance on the publicly available MVTecAD defect dataset.Fourthly,based on the above algorithms,this paper designs and implements a fabric defect detection system that includes user management,image and model management,and fabric defect detection functions.Through testing in a factory setting,the detection speed can reach 2 frames/s.Compared to manual human eye inspection,this system helps workers perform more efficient fabric defect detection.