| Salient object detection is one of the research hotspots in computer vision.Salient object detection based on deep learning is developed from traditional salient object detection methods.However,with the continuous development of salient object detection methods based on deep learning,problems such as gradient dispersion or gradient explosion are gradually found.In this paper,two kinds of salient object detection algorithms based on depth visual perception learning are proposed to solve the problems of inadequate global and local information extraction in salient object detection at the present stage,and the mixed global and local features cannot well express salient objects in images.In order to solve the problem that it takes too much time and cost to use technicians to detect steel surface directly in industrial practice,the proposed algorithm is further applied to strip surface defect detection,and a strip surface defect detection software based on salient object detection is designed.The main research work of this paper is as follows*:(1)Aiming at the problem that most salient object detection algorithms based on deep learning cannot accurately extract global and local salient information fully,this paper introduced feature extraction attention module and multi-scale interaction module on the basis of BASNet[34].A Salient object detection model based on multi-scale feature perception(MSFP)is proposed.Firstly,a feature extraction module is added to extract features from the input image.By extracting features of different scales from the same image and then calibrating them,the feature channels with more noise information are suppressed.Secondly,a feature connection module is added to enlarge the receptive field without losing the feature resolution.Finally,a multi-scale interaction module is added to the decoder to interact the high and low resolution features of different scales,obtain different scale information under different channels,and obtain the saliency object prediction image with smoother and clearer pixels.The main and objective experiments on four common reference data sets show that the proposed algorithm has better detection accuracy than the current mainstream algorithms,more accurate detection of salient objects,and solves the problem of insufficient global and local information extraction of salient objects to a certain extent.(2)Aiming at the low feature resolution of the global features extracted in the above work,based on the work of Zhao[45] et al,Resnet-50 feature extraction network and multi-scale interaction module are introduced in this paper.A salient object detection algorithm based on global perception(GPNet)is proposed.Firstly,feature extraction of input image is carried out by Resnet-50 network.Then,the extracted low-level features are input into the spatial attention module,which is used to extract the foreground area of the image and generate effective low-level features for salient prediction.Secondly,the high-level features are input into the multi-scale interaction module to establish the dependence relationship between the multi-scale channel attention.Finally,the performance of the model is improved and a clearer salient graph is obtained.The main and objective experiments on four commonly used reference data sets show that the proposed algorithm has better detection accuracy than the current mainstream algorithms,and effectively solves the problem of low feature resolution among global features through multi-scale feature interaction,which increases the accuracy of salient object detection.(3)In the detection of strip steel surface defects,the manual detection method has such problems as low detection rate,low accuracy,poor real-time performance,low efficiency,high labor intensity,and great influence of manual experience and subjective factors.A steel surface defect detection software based on salient object detection(SSDD)is designed.First,the user login interface was built,the user name and password were read in and the user identity was verified.Then,the defect detection data set was imported into the database for database association.Finally,the salient object detection algorithm(MSFP)based on multi-scale feature sensing proposed in Chapter 3 was used in strip surface defect detection.Subjective experiments in NEU surface defect database show that MSFP is more effective than the other two common salient object detection algorithms in single-hole strip surface defect detection. |
