Visual Inspection Method Of Fabric Defects Based On Lightweight Deep Neural Networks

In order to meet people's pursuit of high-quality life,textile manufacturers have taken quality control as an indispensable link.At the same time,with the development of science and technology,industrial production methods have developed in the direction of intelligence.Detecting the existence of defects in textiles is one of the important links in controlling its quality.The realization of automatic detection of textile defects is of great significance to ensuring the quality of textile production and improving the intelligence of production links.Most of the traditional fabric defect detection methods have good detection effects on a certain type of fabric,but it is difficult to adapt to a variety of defects and fabric types.The application of deep learning to fabric defect detection has shown good advantages in detecting irregular defect shapes and complex background fabrics.However,the existing algorithm models are usually designed too large,and if deployed on hardware devices,the configuration is low,which is not conducive to meeting the requirements of some industrial tasks for detection speed.In response to these problems,this paper based on deep learning algorithms,proposes a lightweight model fabric defect detection method based on SSD,which improves the detection speed of the algorithm while ensuring the detection accuracy of the algorithm.The main contents of this paper are as follows:(1)Build a fabric data collection platform and use an adaptive correction algorithm based on the gamma function to solve the problem of uneven lighting in the collected data.Through the data enhancement method,the data set is expanded by 9 times and the defect position of the expanded data is automatically marked.(2)A lightweight fabric defect detection method is proposed.The lightweight convolutional neural network MobileNetV3-Large is used as the feature extraction network of SSD,and the target detection algorithm MobileNetV3-SSD is constructed.In order to further improve the detection accuracy,especially for small-area fabric defects,a feature fusion module and an improved RFB module are added on the basis of MobileNetV3-SSD to enrich the semantic information of the shallow feature layer of the network and increase the receptive field,and enhance the feature extraction of the network ability.In order to solve the problem that the position regression loss function is not equivalent to the evaluation and detection index,the loss function is optimized through DIoU to improve the convergence speed and accuracy of the algorithm.(3)The improved algorithm was applied to the fabric defect data set.In order to verify the performance of the improved algorithm,a comparative experiment was designed under the same experimental platform conditions.The results show that due to the efficient network structure design of MobileNetV3 and the optimization of MobileNetV3-SSD,the average accuracy of the improved algorithm on the test set is 84.6%,the detection speed is118 FPS,and the model size is reduced to about one-fifth of the original SSD.This achieves an efficient balance between detection accuracy and speed and makes the algorithm deployment to configuration more efficient.It becomes possible in low hardware devices.(4)Using PyQt as the platform,combined with the improved algorithm proposed to design a fabric defect detection system,realize the real-time display of fabric defect detection results on the system interface.The system includes an online detection module and an offline detection module and has the function of importing the detection result information into the database synchronously.Based on the SSD algorithm,this paper has carried out lightweight processing and improvement,and proposed a lightweight fabric defect detection method.Experiments verify the efficient balance between the detection accuracy and speed of the algorithm,and use it with a software system,it can achieve the purpose of assisting cloth inspection.