A Two-stage Image Edge Detection Algorithm Based On The Coarse-to-fine Strategy And The Concept Of Adjacent Edges

Edge detection is a basic task in image processing and computer vision.It has important application prospects and values in image recognition,target detection,image segmentation,and medical image analysis.Its goal is to identify significant differences in visual perception from the many pixel differences in natural images;the key and the difficult point is to preserve the main structure of the image without the interference of complex textures.Although many specific and representative algorithms have been proposed in the research process of edge detection,there are still some shortcomings.For example,the existing deep learning-based algorithms predict the generally low quality of edges,blurry edges,and existence near edges.A large number of false edges lead to thick edges,which rely heavily on non-maximum suppression post-processing.This paper conducts in-depth research on edge detection algorithms from the following two aspects:1.To solve the problem of low edge quality caused by the mismatch between the existing network structure and the edge detection task,this paper proposes a two-stage image edge detection network algorithm based on a coarse-to-fine strategy.The main contributions are as follows:(1)Through data analysis,we find that the classification difficulty of non-edge pixels in different locations of the image is different,and the closer the image is to the edge,the more difficult it is to classify.Based on this,a two-stage edge detection idea is proposed,which thins the traditional two-class edge detection task into three tasks: edge point,near edge point,and away from edge point.Based on this idea,we design a two-stage W-Net edge detection network from coarse to fine and a multi-weight cross-entropy loss function combined with it.The experimental results show that the ODS index of our algorithm on the BSDS500 dataset is 0.835,which is much better than other algorithms.The edge quality of prediction results has been greatly improved,the edge thinning is very high,and it has good generalization ability.(2)Through observation,it can be found that there are a large number of swing points near the predicted edge.Although these swing points will not affect the index,they will cause the visualization of the prediction result to deteriorate.Based on this,this paper proposes an unsupervised information entropy loss function,which is used to suppress swing points and improve the visualization effect and image quality of the prediction results.The experimental results show that the ODS index on the BSDS500 data set rises from 0.835 to 0.841,the visualization effect of the image has been significantly improved,and the number of swing points is greatly reduced.(3)This paper presents a new label processing method for the BSDS500 dataset.Unlike the existing method,which directly identifies results labeled by more than two labelers as edge points,we think the results labeled by each labeler are valuable,so we use the results labeled by more than two labelers as a benchmark to correct the labeling results of each labeler.Experiments show that our method improves the ODS index of BSDS500 from 0.823 to 0.835 and the OIS index from 0.825 to 0.847 compared with the label processing method of the DSCD algorithm,which improves the algorithm performance.2.To solve the problem of incorrect edge modeling caused by the inaccurate understanding of edges in the existing research and the widespread mislabeling and labeling offset in existing manually labeled edge data sets,this paper proposes a two-stage approach based on a coarseto-fine strategy The main contributions of the image edge detection network algorithm are as follows:(1)In order to solve the problem of inaccurate recognition of edges,this paper proposes the concept of adjacent edges.Edge pixels are composed of pixels of different types of textures that are in contact with each other on the outermost layer.This contact will not be extended to eight neighborhoods.Two different types of non-edge pixels are in contact.We further propose adjacency constraints on the basis of adjacency edges to improve the continuity and edge quality of the edges.Experiments show that the use of adjacent edges can effectively improve the generalization ability of edge detection.When testing the BSDS500 data set across data sets,the ODS reaches 0.736,and the image quality,edge continuity,and edge refinement of the prediction results are greatly improved.It is more sensitive to texture and can well suppress the interference caused by texture lines.(2)To solve the problem of errors in the existing artificially labeled edge data sets,this paper uses a computer to synthesize the adjacent texture synthesis data set based on the proposed adjacent edges.Since the data set is completely synthesized by the computer,there is no Due to human subjective assumptions and other factors caused by edge label offset,missing,wrong labeling,and other issues,the accuracy of the training data is guaranteed.Experimental data shows that the ODS index of the algorithm tested on the texture synthesis data set reached0.978,PSNR reached 28.529 d B,and SSIM reached 0.990,which far exceeded the indexes of other algorithms on this data set.It also passed user surveys and tested across data.In the case of,our prediction results are more popular with users,and the user acceptance rate reaches70.7%.In addition,for the comparison between the test results of the algorithm in this paper on data sets such as BSDS500 and manual annotations,half of the users believe that our results are more accurate than manual annotations.In summary,this data set can completely replace the existing manually labeled data set.