Pixel-patch combination loss for refined edge detection

As a fundamental image characteristic, edge features encapsulate a wealth of information, serving as a crucial foundation in image segmentation networks for accurately delineating and partitioning object edges. Convolutional neural networks (CNNs) have gained prominence recently, finding extensive utility in edge detection. Previous methods primarily emphasized edge prediction accuracy, ignoring edge refinement. In this work, we introduce a novel encoder-decoder architecture that effectively harnesses hierarchical features. By extending the decoder horizontally, we progressively enhance resolution to preserve intricate details from the original image, thereby producing sharp edges. Additionally, we propose a novel loss function named the Pixel-Patch Combination Loss (P²CL), which employs distinct detection strategies in edge and non-edge regions to bolster network accuracy and yield crisp edges. Furthermore, considering the practicality of the algorithm, our method strikes a fine balance between accuracy and model size. It delivers precise and sharp edges while ensuring efficient model operation, thereby laying a robust foundation for advancements deployed on mobile devices or embedded systems. Our method was evaluated on three publicly available datasets, including BSDS500, Multicue, and BIPED. The experimental results show the superiority of our approach, achieving a competitive ODS F-score of 0.832 on the BSDS500 benchmark and significantly enhancing edge detection accuracy.