Feature enhancement: predict more detailed and crisper edges

Abstract

CNN-based methods have improved the performance of edge detection in recent years. While a common issue with the most recent methods is that there is a thickness problem in predicting the edge map, and when objects are small or have dense edges, the predicted edge lines are blurred. We know that multi-pooling reduces the resolution of features, and using a balanced cross-entropy loss function will also make the predicted edges thicker. In this paper, we propose a multi-scale feature hybrid network for edge detection to improve edge resolution. Multiple long and short residual connections are set to establish various information propagation pathways. We construct a feature enhancement unit in the up-sampling path of the network to obtain multi-scale features and fuse higher-resolution features. We demonstrate that residual connections in the network can overcome the class-imbalance training problem in edge samples. Moreover, we introduce a new biased cross-entropy loss function to accomplish the training of our network better, which adjusts the weights according to the ratio of edges and non-edges pixels. Compared to other methods, our network can predict clearer and sharper edges with more details. Evaluate the network on BSDS500 and NYUDv2, our method achieves ODS F-measure of 0.832 on the BSDS500 dataset and 0.768 on the NYUD dataset, best than current state-of-the-art results.