Research On Plain Fabric Defect Detection Algorithm Based On Convolutional Neural Network

The weaving industry has been an important pillar of China's economy and it occupies an irreplaceable position in the production and social life of the national economy.The detection and identification of fiber defects is an important factor restricting the production efficiency and quality of the textile industry.Traditional fabric defect detection and recognition are performed manually,and there are problems such as slow detection speed^[1],high labor intensiveness,and the influence of inspector experience.At present,the algorithms used for automatic defect detection and recognition of cloth usually use the extraction of easily quantified features to achieve the purpose of defect detection,but the defects with the characteristics of small area ratio,large scale change,and complex shape are difficult to extract and identify well,and finally lead to The limitations of the types of defects identified,At the same time,with the development of industrial production rate and fabric quality standard,the accuracy and speed of detecting fabric defects must be improved.Deep learning neural network algorithms are widely used in image recognition scenes.convolutional neural system cloth defect detection and recognition algorithm proposal to increase Thus,the possible defects in the treatise is so that the efficiency of the algorithm is further improved.First of all,the research object of this paper is plain coloured cloth defect image,through to the fabric defect data set 20 kinds of defects of data statistics and visualization analysis,aiming at the problem of unbalanced defect data of using flip vertical,horizontal rotation data enhancement method.By comparing the experimental results of four kinds of commonly used network models,the selection of feature extraction for Resnet50 network are established the first generation of testing model,after the experiment was tentatively benchmark data accuracy,this model is the cloth defects classification,location and function.Secondly,in view of the defect area accounted for a small problem,using feature fusion network,enhancement of tiny flaw detection ability;To solve the problem of the complex shape of cloth defects,two layers of Deformable Convolution Networks(DCN)were added in the last stage of Resnet50 to adapt to the defect contour of different shapes.Experiments showed that according to the characteristics of the data set using two kinds of improved algorithm accuracy was improved after the modification of neural network structure.Then,in view of the defect area the problem of large scale changes,the first to use K-means clustering algorithm(K means clustering algorithm)get cloth defects distribution regularity of regional scale,and then to the aggregation of 9 kinds of defective area aspect ratio,The experiment results show that the model can effectively improve the accuracy and precision in the detection of plain cloth defects,and the detection speed can reach 85ms,which can meet the requirements of factory for cloth defects detection.Finally,summarizes the full text work,gives several can continue in the future research direction and application field...