The multi-scale channel attention full connection network for edge detection
Meng Li, Wei Wu · Journal of Physics Conference Series · 2021
Abstract Edge detection algorithms based on deep learning have made great progress due to its super feature extraction and presentation ability in recent years. However, existing edge detection algorithms encounter some problems that image features extracted by networks are not sufficient, and the generated edges are fuzzy with large gap compared to the ground truth (GT). To solve the abovementioned problems, we have proposed a multi-scale channel attention full connection network model for edge detection (MAED) in this work. The proposed network consists of three parts: multi-level feature extraction part, multi-scale deep supervise part and fusion part. In addition, by introducing the "attention" mechanism into the networks, underlying features of the input image can be fully extracted and the training convergence speed can be accelerated simultaneously. To evaluate the performance of the proposed algorithm, we conducted a comprehensive study on some commonly-used datasets. The final experimental results demonstrated that our network achieves excellent performance based on both qualitative and quantitative analysis.