Edge Detection based-on U-Net using Edge Classification CNN

Edge detection is the first necessary step in image processing for object segmentation, detection, and recognition. The Canny algorithm is widely used filter-based approach, but it requires the correct adjustment of its parameters according to the variations in images. In this paper, we propose a method that is consisted of two steps for the robust detection of edges in an image. The proposed algorithm adopts convolutional neural networks that can handle the diverse variations caused by illumination, pose, and scale change. First, we train a convolutional neural network to decide whether a given input edge image is good or not. We can generate as many training images as we want using this network. Finally, U-Net is used to generate an edge image using a gray image as input. Experimental results show the robustness of the proposed algorithm for images acquired under outdoor and indoor environments.