Abstract:

Most of the deep-learning based dehazing models have fixed receptive filed after the parameter are fixed. As a result, the dehazing network cannot adopt the optimal mode for dehazing each specific scene, resulting in ambiguity and distortion in the results. In view of these problems, this study proposes a dynamic receptive field feature selection dehazing network. A feature-attention atrous block with atrous convolution is designed as the basic module of the network. Multiple feature attention atrous blocks with different atrous rates are used in parallel to extract multi-scale features. Dynamic feature fusion is performed on these features to form a dynamic receptive field block. Multiple dynamic receptive field blocks are combined with residual connections to form a deep network. The features from different levels are dynamically mixed and decoded to obtain a haze-free image. The experimental results show that the proposed algorithm has a good dehazing performance on indoor, outdoor, and real hazy images, and can generate clear and natural dehazing images.

Key words: image dehazing, dynamic receptive field, multi-scale features, dynamic feature fusion, atrous convolution, self-attention mechanism, dynamic neural network, dynamic parameters