| High-quality textiles not only require the fabric to be strong and durable,but also require the surface of the fabric to be smooth and free of defects.Therefore,in the process of textile production,product quality inspection is a very important link.Timely detection and repair of defects can greatly reduce the loss of enterprises.With the increase of fabric types,the types of defects have become diverse.At present,fabric defects generally include broken warp,broken weft,hole,yarn jump,knot,stain and so on.To ensure the efficacy of fabric defect detection,manual detection must be supplanted with automated technology,which would not only enhance production efficiency but also enhance fabric quality.Manual detection is a traditional method,yet it has a low detection speed,high missed detection rate and hefty cost,making it inadequate for real-time and precise detection in industrial production.With the popularity of artificial intelligence,the traditional extraction methods based on eigenvalues and low-dimensional pixel features are gradually replaced by data-based intelligent learning algorithms.Heuristic learning algorithms,in comparison to traditional ones,possess the benefit of high recognition accuracy,strong generalization capability,no requirement for intricate analytical relations,and a minimal sensitivity to hyperparameters.In recent years,computer vision has seen tremendous advances in object detection and classification,with the algorithm based on convolutional neural network model being a novel research direction.This model utilizes multiple nonlinear transformation training,which is capable of extracting features from images at multiple levels and can effectively identify the intricate textures of fabric images.The model has strong applicability.Consequently,the utilization of a convolutional neural network algorithm for fabric defect detection can be successful in completing the task and significantly augmenting the speed and precision of detection.This thesis delves into the convolutional neural network-based fabric defect detection algorithm,optimizing and enhancing its structure by incorporating texture characteristics of fabric defects.This model not only increases accuracy in detecting fabric defects,but also reduces model parameters and hastens detection.In this thesis,a Mesh-Faster RCNN model is proposed,which replaces the RPN layer in the original Faster RCNN model with × mesh.Instead of using anchor in the RPN layer to match a single target,it is improved to use mesh,that is,more matches may fall into all targets in the mesh,and less candidate regions are generated.It reduces the burden of subsequent data transmission and processing,thus greatly shortens the memory consumption of the computer,saves a lot of time,and greatly accelerates the efficiency of detection.By applying Logistic Regression algorithm,we can identify and locate each kind of defect more effectively.Moreover,compared with the traditional Faster RCNN model,this method can effectively reduce the parameters,thus greatly improving the efficiency of detection. |
