Lightweight Edge Detection Model with Parallel Pixel Difference

Although existing learning-based edge detection methods have made significant progress, they typically require large network parameters and substantial computational resources, which limits their practical applicability. Instead of focusing on larger deep models to improve detection performance, we propose a lightweight deep edge detection method that reduces computational costs while achieving comparable detection performance. This approach is built on a residual framework network architecture. Specifically, we introduce an efficient Parallel Pixel Difference Convolution Block (PPDB) as the core building block of our method to reduce model complexity. Additionally, to better leverage features from different scales, we incorporate a lightweight dynamic upsampling technique and develop a Refined Feature Fusion Block (RFFB) to achieve high-quality results. Experimental results on benchmark datasets show that our method requires about 200K fewer parameters and FLOPs compared to state-of-the-art methods such as LDC, while maintaining competitive performance.