Multiscale hierarchical attention fusion network for edge detection

Edge detection is one of the basic challenges in the field of computer vision. The results of most recent methods produce thick edges and background interference. The images generated by these networks must be postprocessed with non-maximum suppression (NMS). To tackle the problem, we propose a novel edge detection model that allows the network to concentrate on learning the contextual features of an image, thereby obtaining more accurate pixel edges. To obtain abundant multi-granularity features of image high-level features, we introduce multi-scale feature stratification module (MFM). Then, we increase the constraint between pixels through the edge attention module (EAM), so that the model can obtain stronger feature extraction ability. These new approaches can improve the ability of describing edges of models. Evaluating our method on two popular benchmark datasets, the edge image predicted by this method is superior to existing edge detection methods in subjective perception and objective evaluation indexes.