Multi-decoding Network with Attention Learning for Edge Detection

Xiao Zhang, Chuan Lin

Abstract: In the past few years, edge detection models based on convolutional neural networks have made remarkable progress. These models consist of the encoding network and the decoding network. The classification networks (e.g. VGG16) are generally used as the encoding network and the researchers mainly focus on the design of the decoding network. Because natural images contain many edges with different scales, how to make full use of rich and hierarchical convolutional features for edge detection is pivotal. We propose a decoding network that effectively integrates multi-level features from all convolutional layers, named multi-decoding network. Firstly, the side outputs from VGG16 are divided into shallow-, mid-, and deep-level edge features. These edge features are processed separately by multiple independent decoding networks composed of several refinement blocks. Subsequently, to improve the ability of edge fusion, we design a multi-level top-down architecture and a multi-scale attention module, which utilizes the attention mechanism to gradually refine edges. Ultimately, three types of predictions are averaged to produce the final prediction. We evaluate our method on the BSDS500, NYUDv2, and Multicue datasets. Experimental results demonstrate that our model is superior to multiple state-of-the-art edge detection models.