Hierarchical contrastive learning and color standardization for single image sand-dust removal

Abstract

Convolutional neural networks (CNN) have demonstrated impressive performance in reconstructing images in challenging environments. However, there is still a blank in the field of CNN-based sandstorm image processing. Existing sandstorm removal algorithms enhance degraded images by using prior knowledge, but often fail to address the issues of color cast, low contrast, and poor recognizability. To bridge the gap, we present a novel end-to-end sand-dust reconstruction network and incorporate hierarchical contrastive regularization and color constraint in the network. Based on contrastive learning, the hierarchical contrastive regularization reconstructs the sand-free image by pulling it closer to ’positive’ pairs while pushing it away from ’negative’ pairs in representation space. Furthermore, considering the specific characteristics of sandstorm images, we introduce the color constraint term as a sub-loss function to balance the hue, saturation, and value of the reconstructed image. Experimental results show that the proposed SdR-Net outperforms state-of-the-arts in both quantitative and qualitative.