PiDiNeXt: An Efficient Edge Detector Based on Parallel Pixel Difference Networks

The Pixel Difference Network (PiDiNet) is well-known for its success in edge detection. Combining traditional operators with deep learning, PiDiNet achieves competitive results with fewer parameters. However, the complex and inefficient choice of traditional edge detection operators hinders PiDiNet's further development. Therefore, we propose a novel lightweight edge detector called PiDiNeXt, which combines traditional edge detection operators with deep learning-based model in parallel to solve the operators choice problem and further enrich features. The results of experiments on BSDS500 and BIPED datasets demonstrate that PiDiNeXt outperforms PiDiNet in terms of accuracy. Moreover, we employ the reparameterization technique to prevent the extra computational cost caused by the multi-branch construction. This enables PiDiNeXt to achieve an inference speed of 80 FPS, comparable to that of PiDiNet. Furthermore, the lightweight version of PiDiNeXt can achieve an inference speed of over 200 FPS, meeting the needs of most real-time applications. The source code is available at https://github.com/Li-yachuan/PiDiNeXt.