Research On Fabric Defect Detection Algorithm Based On Deep Learning

China is the world's largest fabric producer,defect identification and inspection of fabrics are important factors that restrict its production efficiency and product quality.The traditional cloth defect detection and identification is carried out manually,and there are problems such as subjective influence and low detection efficiency.In recent years,with the breakthrough of computer computing power,deep learning technology has developed rapidly,and more and more applications have been produced in the field of defect detection in industrial production.Cloth defect detection should also be combined with deep learning to achieve better recognition results and detection accuracy.Different from common image recognition and detection scenes,cloth defect images have the characteristics of single background,repetitive textures,large proportion of small target defects,and extreme aspect ratio defects.Therefore,this thesis studies a cloth defect detection algorithm based on deep learning.The main contents of this thesis are as follows:1.A two-step pooling algorithm is proposed.For the single background and repetitive texture of the defective image of the cloth,the two-step pooling algorithm can suppress the background of the feature map and enhance the characteristics of the defective point.The two-step pooling algorithm is divided into three steps: first,the global average pooling is used to obtain the background approximation of the shallow feature map;then the background of the feature map is suppressed by the designed background suppression function,and the defect point features are enhanced;and finally maximum pooling to extract the features of the defect area.Due to the simple structure of the two-step pooling algorithm,this thesis combines it with three types of classic feature extraction networks and conducts related control experiments on the DAGM data set.The results show that the two-step pooling algorithm can greatly improve the recognition effect of defects with a small difference from the background.2.Improved the general object detection model Faster R-CNN.Aiming at the difficulty of fabric defects with large proportion of small target defects and extreme aspect ratio defects,an improved multi-scale detection scheme I-FPN was added to the Faster R-CNN detection framework.I-FPN mainly improves the FPN upsampling method,and adds a bottom-up path to enhance the fusion effect of the top layer and non-adjacent layer features.Finally,the Anchor size is set by clustering.At the same time,a two-step pooling algorithm is added to the backbone network of the detection framework.This thesis compares the algorithm model before and after the improvement on the self-collected cloth defect data set.The results show that compared with the original Faster R-CNN,the detection model with the multi-scale detection scheme has a certain improvement effect on the detection of small targets.Compared with FPN,I-FPN has a certain improvement in detection accuracy,and has a better detection effect for extreme aspect ratio defects.The detection model with a two-step pooling algorithm has better detection capabilities for defect types with a small difference from the background.3.Designed and implemented a software system for fabric defect detection.This system can detect a batch of cloth images and output a test report,so that quality inspection personnel can efficiently complete the test task.