Learning Multi-Scale Features using Dilated Convolution for Contour Detection

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