Research On Visual Detection Algorithm Of Periodic Texture Surface Defects

Machine vision technology has been widely applied in the quality inspection of various industrial products in recent years,to replace inefficient manual visual inspection methods.With the continuous expansion of scale,the number of industrial products with periodic texture surfaces is increasing rapidly.How to effectively detect surface defects of periodic texture products has become an important research direction in the field of machine vision.This paper studies the above problems,and the main work includes the following aspects:Establishment of a fabric surface defect dataset.Considering the influence of various factors such as lightness,flatness,and camera resolution in the real-time detection process,the detection accuracy of the algorithm is not able to achieve the expectations.In order to facilitate the follow-up research experiment,we sampled and labelled an image dataset of solid color,stripe,and lattice texture: fabric-defected.Then the images are pre-processed by gray-scale processing and sub-image segmentation,which is convenient for the subsequent verification research.Aiming at the characteristics of target detection tasks for surface defects,this paper proposes a defect detection model based on improved Faster R-CNN.In the algorithm,residual network is used as the main structure for feature extraction,to maintain the original defect information of periodic texture images by utilizing the universality of object detection.Due to the Region Proposals Network structure of Faster R-CNN can effectively generate prior boxes,the method is effective.Then the model is trained based on different backbone networks.Aiming at the periodic characteristics of surface texture,a detection algorithm based on Multi-metric-Multi-model Image Voting is proposed.In the experiment,the periodic texture statistics feature is used to select the method of sub-image segmentation.Firstly,the characteristic values of the multi-dimensional change level statistical measures of the subimages are sorted out and calculated,and then the standard value of the flawless sub-image background is extracted by Zero-Slope-Random Sampling Consensus method.After that,the statistical characteristic values and the standard value are passed to the scoring voting models for sorting,and the sub-image labels belonging to the defective foreground and background texture are judged.The research results demonstrate that due to the high fusion degree of the defect foreground and background in the non-large object dataset,and the large differences between the classes of textile defects,the loss value calculation of the self-sampling bounding box will fluctuate in the process of model classifier regression,so that the detection algorithm based on improved Faster R-CNN has the highest detection accuracy of 39.8% after multiple rounds of experiments.There is still room for improvement in the final detection effect.In the Multi-metric-Multi-model Image Voting,the gray mean measure and Borda voting model performed the best,the prediction accuracy reached 95.6% at confidence 0.25,the average time and space also achieved fair results.