Assessment of Deep Learning Models for Image Edge Detection

Edge detection is a crucial technique in image processing, essential for various applications, including feature extraction, object identification, and segmentation. Conventional methods, such as Sobel and Canny filters, are not suitable for complex image structures and different lighting conditions in the image. Deep learning-based edge detection algorithms perform better in these situations and become an essential component of edge detection. ‎ This study presented a deep learning-based approach for detecting edges in objects and complex backgrounds using the U-Net and its variants. This study used the BSDS500 dataset to compare eight deep learning-based edge detection algorithms for object and background out-‎lines. The testing outcomes show that the residual structure based on Resnet50-Unet outperformed in detecting the outlines of objects and backgrounds, achieving 95% accuracy, 0.62 precision, 0.89 recall, and 0.71 F1 score without augmentation of the BSDS500 dataset. ‎The performance of Resnet50-Unet is enhanced when the augmented BSDS500 dataset is utilised, and it achieved 97% accuracy, 0.72 precision, 0.98 recall, and 0.83 F1 score. The results obtained are also cross-validated using the BIPED dataset. The proposed algorithm efficiently detects edges and object boundaries in real time.