Research On Automatic Detection Algorithm Of Cotton Flaw Image

Fabric products such as cotton cloth have always been an indispensable part of our people's daily life,but some defects may appear in the manufacturing process of cloth,which will seriously affect the performance and quality of its products.In the current cotton production enterprises,there are still a lot of manual detection work,wrong inspection,omission inspection and other problems cost much manpower,financial resources and time,and the efficiency is very low.Therefore,automatic detection of cotton defects based on image processing is widely concerned by enterprises,and how to accurately and efficiently detect cotton defects has become one of the research hotspots in this field in recent years.With the development of artificial intelligence,the traditional cotton defect detection method urgently needs to realize intelligent transformation.This thesis,which is an important part of automatic detection system for cotton defects,mainly studies the cotton defect detection algorithm in the aspect of target detection in image processing and deep learning for the demand of cotton defect detection.In order to solve the problems of high cost,low accuracy and slow speed of defect detection in traditional cotton production process,a FS-YOLOV3(Four Scales YOLOV3)network is proposed to automatically detect cotton defects.According to the characteristics of cotton defect image,the main work is as follows:(1)In image processing algorithm,this thesis designed the preprocessing steps,using the enhanced algorithm(linear transformation enhancement,contrast adaptive histogram equalization,frequency domain filtering)increase the background and the defects of contrast,using noise processing algorithm(morphological method,median filtering,gaussian filter)to reduce the influence of irrelevant factors for accuracy.In this thesis,Gabor filtering based on feature extraction,Sobel operator detection based on edge information detection,Canny operator detection based on support vector machine defect classification and contour search methods are studied,in order to reduce noise information,optimize the model,increase the accuracy of image feature extraction,improve the robustness of the algorithm.(2)In deep learning algorithm,based on a stage of target detection are studied YOLOv3 algorithm,based on the two-phase Faster R-CNN algorithm of target detection,and improvements on YOLOv3 algorithm.This thesis is designing the FS-YOLOv3algorithm: through the design of four different scales of convolution characteristic figure,batch normalization(Batch Normalization,BN),the global average pooling(GAP)Global Average Pooling,optimize network structure;Combined with K-means++ clustering algorithm,theanchor frame with better size can be obtained to improve the detection speed.The Softer NMS algorithm as the prediction filtering mechanism,makes the high classification confidence border position more accurate;The coefficient is added to the loss function to make the detection accuracy of small targets higher.Through experimental comparison,FS-YOLOv3 algorithm has improved in accuracy,recall rate and accuracy.This thesis is based on Windows platform,Python language,Tensor Flow deep learning framework,Py Charm as the development platform,with open source data set AITEX for testing,and designed an automatic detection system for cotton defects,which is simple in operation,economical in cost,and improved detection efficiency.Through the comparison of different algorithms and a large number of sample tests,it can be known that the FS-YOLOV3 algorithm proposed in this thesis,it can effectively improve the detection accuracy of cotton defects for low-contrast and small-scale targets,and its overall performance is better than that of traditional detection methods.The detection system can effectively reduce the influence of defects on the quality of cotton cloth,improve production efficiency,and has a good application value.