Focusing on Refinement in Edge Detection Using Detailed Contour Loss

Edge features, as a fundamental aspect of images, contain a wealth of information and serve as a critical basis for accurately defining and segmenting object boundaries in image segmentation networks. Previous approaches primarily focused on achieving high accuracy in edge prediction while often overlooking the aspect of edge refinement. In this study, we propose a Horizontally Extended Refinement Network (HERN) that effectively leverages layered features and progressively enhances resolution to retain detailed elements from the original image, leading to well-defined edges. Furthermore, we introduce an innovative loss function called Detailed Contour Loss (DCL), which incorporates the image's geometric details to enhance network precision and produce distinct contours. Additionally, our method achieves an optimal balance between precision and model size, ensuring accurate and crisp edges without compromising operational efficiency. This strategy provides a robust groundwork for future developments in mobile and embedded systems. Our method is tested on BSDS500 and BIPED datasets, where it achieves an ODS F-measure of 0.826, surpassing current leading methods on BSDS500.