RSTNet: sandstorm image enhancement using a residual saturation and transmission estimation network

Abstract The degradation of sandstorm images is diverse, as it encompasses color casting and dark and dim conditions. In addition, because the structural degradation of sandstorm images involves imbalanced color channels, an image may seem yellowish or reddish. Moreover, when an image is obtained, because the atmosphere contains various hazy particles such as sand, the sand in the image may seem hazy. The existing sandstorm image enhancement methods improve the image’s color components and mitigate the hazy effect. However, an artificial color component may appear when the enhanced color channels are imbalanced. Therefore, in this paper, an effective color balancing method using convolutional neural networks and image saturation is proposed because image saturation is strongly related to color information. Additionally, a color-balanced image may be hazy because of the presence of hazy particles; thus, in this paper, hazy images are enhanced using transmission networks for dehazing. Overall, a sandstorm image enhancement method using a convolutional neural network that estimates residual saturation and a transmission estimation network (RST-Net) is proposed. The enhanced images obtained using the RSTNet appear natural, and the RSTNet is verified to be suitable for enhancing degraded sandstorm images.