Multi-scale Haze Removal via Residual Network

The scattering of climatic particles significantly alters images captured under unfavorable weather situations. Although many traditional methods have efficiently committed to annihilating haze, they pose some limitation due to their hand-crafted features, e.g., dark channel, maximum contrast, and colour disparity. This paper presents an end-to-end algorithm to restore a hazy image using a residual-based deep CNN straightforwardly. The proposed algorithm is non-subject to the climatic dispersing model, yet it learns the mapping relationship within the hazy input image and their corresponding transmission map. The network architecture constitutes a convolution kernel and multi-scale fusion layers in extracting relevant features in predicting a holistic propagation map. Ultimately, we obtain the residual image and circumvent the loss of information using a residual network. Comprehensive empirical results demonstrate that the proposed technique outperforms multiple conventional algorithms.