Learning Multi-Scale Features Using Dilated Convolution for Contour Detection

For the contour detection task, we use the EfficientNet model as the backbone network and propose a network model that uses dilated convolution for multi-scale optimization. The network is accumulated top-down layer by layer, combining multiple optimization modules concat together to achieve a richer feature representation. To fuse feature information at different scales, we introduce a new Multi-scale optimization module to replace the use of deeper network structures or more complex decoding methods, which uses channel attention module to learn the correlation between channels and then uses dilated convolution of different scales to enhance contextual information. High generalization performance and accuracy are obtained in comparison with recent deep learning-based contour detection models. We evaluate our approach on two datasets, i.e., BSDS500 and NYUD-v2, achieving an ODS F-measure value of 0.828 on BSDS500. In particular, the results of BSDS500 exceed the human-level performance under more stringent criteria.