Research On Saliency Detection Method For Strip Steel Surface Defects

As an important metal plate and strip in our country,when there are defects on the surface of strip steel,it will not only affect the aesthetics and comfort of the product,but also will have a bad impact on its performance and safety.However,the current defect detection method utilizing manual inspection is not enough to meet the requirement of current production automation.That urgently requires an efficient online quality inspection system aim to promote improving the surface quality of the products,and now the surface defect image detection technology based on machine vision has become one of the hot topics of product quality control.In order to quickly find useful information from plenty of images,this paper introduces visual saliency detection technology.This technology can simulate the human visual attention mechanism and will choose the important visual parts to be processed in priority when receiving scene information.Therefore,the introduction of visual saliency detection method into the surface defect detection of strip steel can greatly improve the efficiency of defect detection.In addition,the result of introducing visual saliency detection is more in line with human visual cognition.However,the experimental results show that the performance of the existing saliency detection methods is not ideal in the defect detection database of strip steel,especially dealing with defect images with cluttered background,low contrast and noise interference.In this paper,in view of the shortcomings of existing saliency detection methods,we carry out in-depth research and abundant experimental verification,and then propose two effective saliency detection algorithms.The main research contents of this paper are as follows:(1)In order to solve the problem of imperfect evaluation criteria in the field of surface defect detection,this paper uses the open annotation tool:LabelMe to annotate three types of defects(including Inclusion,Patches,and Scratches)consisting of total of 900 defect images,and generate the corresponding pixel-wise ground truth.That constructs the benchmark database,called SD-saliency-900,which will be used to evaluate the performance of different saliency detection methods.(2)In this paper,we formulate the surface defect detection of strip steel as a saliency evaluation problem,and propose a saliency detection algorithm based on multiple constraints and improved texture features,i.e.,MCITF.This method constructs 83-dim welldesigned texture feature bank,which efficiently solves the problem of insufficient feature discrimination of existing saliency detection methods.After that,based on the framework of multiple-instance learning,we can obtain reliable label matrix,then we introduce Laplacian regularization and high-level prior constraints.The closed form optimal solution is obtained by solving the model.Thus,we obtain the high-quality saliency map.Experimental results demonstrate that our proposed MCITF model consistently outperforms state-of-the-art detection methods,and has better robustness.(3)Aiming at meet the high requirements of detection accuracy,real-time performance and robustness for online monitoring in the factory at the same time,this paper takes full advantage of deep learning and proposes a saliency detection algorithm based on encoderdecoder residual refinement network,i.e.,EDRNet.This algorithm uses the feature extraction module of ResNet to learn rich feature representation in the encoder stage,and integrates convolution attention mechanism to focus on feature extraction of important regions.In the decoder stage,the channels weighted block and the residual decoder block designed in this paper are used alternatively to effectively integrate the context information,and recover the more complete spatial predicted saliency values step by step.Thus,the rough saliency map can be obtained.Finally,by utilizing the residual refinement structure proposed in this paper,the final refined high-quality saliency map can be obtained,which can efficiently and uniformly highlight the complete defect objects containing with little or even no background noise,and has well-defined object boundary.The detection results are close to or even better than ground truth.Compared with the state-of-the-art saliency detection methods based on deep learning,EDRNet achieves the best detection accuracy and robustness,and the actual detection efficiency is as high as 27fps.