Dynamic Texture Defect Detection Based On Video Sequence Analysis

Defect detection is a very important part in industrial application.In recent years,the algorithms of defect detection based on machine vision are mainly based on various state-of-the-art deep neural network frameworks,and detect defect directly on surface images.However,in practical applications,especially when using robot inspection for defect detection,there will be problems such as dynamic background interference,lack of defect samples,and low speed of direct detection.To address the problems of defect detection in robot inspection video,this thesis researches on algorithms for video object segmentation,video sequence anomaly detection and surface texture defect detection,and implements a fast defect detection software for outdoor robot inspections based on the above algorithm.The main contents are as follows:(1)Aiming at the problem of video defect detection region segmentation,a video object segmentation algorithm based on weak temporal information is designed.Using ResNet network as the basis for video object appearance feature extraction,the algorithm extracts weak temporal information between consecutive frames based on the mask of the previous frame,and integrates different scale and temporal features to improve the accuracy of video object segmentation.On the DAVIS dataset,the contour accuracy and regional similarity reached 87.34%and 80.65%,(2)In view of the low speed of direct detection in video defect detection,a fast anomaly detection algorithm based on prediction model is designed.Utilizing the temporal correlation of non-defective video pixels,a video sequence prediction model based on Attention-LSTM is designed,and based on the prediction error,the video frames with defects in the detection sequence are quickly located.The processing speed of the algorithm based on the prediction model is about 4 times of the detection model,which greatly improves the overall speed of the inspection video defect detection.(3)Implemented texture defect detection algorithms based on deep learning.Research on the state-of-the-art object detection models,including Faster R-CNN,SSD,YOLO-V3,and RetinaNet,and consider the performance of each model on public data sets,choose Faster R-CNN as the framework for defect detection algorithms and further research is performed on the subject data set.The problem of lack of defect samples in practical application is solved through model migration.(4)A texture defect detection software based on video analysis is implemented.In response to the defect detection requirements of machine inspection,a texture defect detection software based on video analysis is developed.The software can segment the video object,quickly locate the defect image in the sequence,and classify defects.