An Enhancement of the U-Net Model: A Deep Learning Framework for Thin and Noise-Free Edge Detection

Edge detection is critical in a variety of real-world scenarios, from medical imaging to autonomous vehicles, robotics, and surveillance. Misdetecting edges due to noise, complex structures, or environmental factors can lead to significant issues, such as misdiagnoses in healthcare, accidents in autonomous driving, or failures in manufacturing and security systems. Thus, accurate, noise-free edge detection is essential for improving human decision-making and operational efficiency in these do- mains. This paper presents a modified model specifically designed to address the challenge of detecting thin, noise-free edges in such scenarios. While existing models like DexiNed and PiDiNet excel in different areas—DexiNed offering high accuracy through its dense multi-scale architecture at the cost of higher computational overhead, and PiDiNet being optimal for real-time edge detection in resource-constrained environments—they both struggle with capturing fine, noise-free edges. U-Net performs better in this aspect, but its overall edge detection performance remains inferior to our proposed model. Our approach demonstrates superior edge detection capabilities, achieving a lower noise probability of 1.847%, and a signal-to-noise ratio (SNR) of 0.44, which surpasses the performance of DexiNed, PiDiNet, and U-Net. These results indicate that our model effectively balances noise reduction and edge preservation, making it highly suitable for detecting fine edges in complex image structures.