Semantic Edge Detection with ConvNeXt and Bi-directional MLA

This paper focuses on semantic edge detection based on physical properties (cause of occurrence), which classifies edges into four classes: reflectance, illumination, normal, and depth edges. This semantic edge detection is widely used in consumer electronics and industrial manufacturing as a complement to computer vision technology. In this paper, a convolutional neural network (CNN)-based network is proposed to improve the performance of physical property-based semantic edge detection. The proposed CNN-based network has advantages in terms of memory and speed compared to the Transformer-based approaches. In addition, this paper analyzes the loss function that is effective for the proposed network. The proposed method shows superior performance compared to the previous state-of-the-art, with improvements in ODS, OIS, and AP of 1.1%, 4.7%, and 2.7%, respectively, and qualitatively demonstrates improved edge detection and classification performance.