Cloth Defect Recognition Based On EfficientDe

In order to better regional segmentation and improve the accuracy of fabric defect recognition,this paper mainly studied fabric defect detection and classification from two aspects.The first point is to study the traditional edge detection algorithm to segment the fabric defect area.In this experiment,the improved Canny edge detection algorithm is used to segment the defect area effectively.There are two optimizations in this paper for the traditional Canny edge detection algorithm.The first point is one of the traditional Canny edge detection algorithm using gaussian filtering approach to noise,but with gaussian filtering denoising ability enhancement,fuzzy edge defect,in operation,threshold lag will get rid of a fuzzy edge of these result in edge detection of defect information incomplete,therefore identified defect is discontinuous even residual defect area.In order to solve this problem,an adaptive bilateral filter with edge preserving function is used to replace gaussian filter for denoising.Experimental results show that the defect edge is clearer after denoising by bilateral filtering algorithm.The second point is that the traditional Canny edge detection algorithm performs hysteresis thresholding operation through artificially set high and low thresholds.Because the result of threshold setting directly determines whether the segmentation region is complete,the way of threshold setting is optimized.The improved algorithm mainly set the high and low thresholds adaptively through the iterative method,so that each defect image had a corresponding high and low threshold.The improved Canny edge detection algorithm was used to detect the target defect area,and the clear fabric defect image was obtained.The second point is to study the application of artificial intelligence and deep learning in the field of defect recognition.EfficientDet is a set of algorithms that the Google team proposed in 2020.Research institutions and research institutes study it because of its small model and high efficiency.The experiment uses the improved EfficientDet algorithm to detect and classify the defects.The improved algorithm takes Efficient Net as the backbone network and adds the optimized Bi FPN module for deeper feature extraction.The pre-training model is used for transfer learning.In order to optimize the gradient stability and avoid the over-fitting problem,swish function is used for normalization.Local Loss function is used to solve the problem of unbalanced positive and negative samples.