Learning Gabor layer for edge detection network

In recent years, lightweight edge detection research has attracted much attention from scholars. This type of model has the advantages of low computational complexity and small parameter scale but sacrifices detection accuracy. The main reason is that the model does not fully extract the underlying and detailed features. We construct a learning Gabor layer as the front part of the edge detection network (“encoding network + decoding network”) to enhance the model's capability to extract low-level and detailed features. In addition, it can assist the network in extracting contextual semantic features and aggregating edges that continue in the same direction. The proposed learning Gabor layer only contains 150 parameters, and the total size of the model parameters is merely 0.49M. Applying the learning Gabor layer to our designed lightweight edge detection network (baseline) and validating it on four benchmark datasets (BSDS-VOC, NYUDv2, BIPED, Multicue) demonstrate the effectiveness of the learning Gabor layer.