Multi-residual Connection Network for Edge Detection

CNN-based methods have improved the performance of edge detection in recent years. However, the edge map predicted by the neural network has a problem of thickness. We analyze that this is due to the training problem caused by the unbalanced distribution of edge and non-edge pixels. We propose to use residual connections to solve the class-imbalanced training problem in edge samples. Moreover, we introduce a new biased cross-entropy loss function to better train the edge detection network, which will adjust the weights according to the ratio of edges and non-edges pixels. Compared to other methods, our method predicts clearer and crisp edges. We set up multiple long and short residual connections in the network to establish various information propagation pathways. This makes the final prediction of our method not only with contour information but also have rich details. We evaluated our method on the BSDS500 and NYUD datasets and showed promising results.