SEHSNet: Stage Enhancement and Hierarchical Supervision Network for edge detection

Edge detection is a challenging low-level vision task. After a long period of development, there are already edge detectors whose edge recognition performance exceeds that of humans. However, there are still three problems in the existing modern deep CNN edge detection methods, which can be summarized as low-level under chaos, high-level over suppression, and inter-layer without interaction. For the first problem, we propose a cascade structure of dilated branches with truncated gradient flow, called the low-level proofreading module. A high-level cognition module is used to excavate the global edge cues to solve the over-suppression problem of high level. Considering the limitations of deep supervision, we create a deep hierarchical supervision strategy to facilitate inter-layer communication. We combine all the components to form our network, Stage Enhancement and Hierarchical Supervision Network. It is from the perspective of the deep edge detector, synthesizing and optimizing the advantages of previous works, and creating a new architecture to accommodate edge detection tasks. Our network outperforms the SOTA algorithms on several metrics. For example, when only using BSDS500 dataset for training, we achieved the F -measure ODS of 0.820, breaking the previous best record of 0.808.