IHR-Net: image haze removal network based on improved YUV spatial strategy

Image haze removal is a challenging task in computer vision. Existing dehazing algorithms are typically applied in the RGB space, assuming that haze affects all color channels equally, and thus enhancing the image independently across the three channels. Actually, compared to the RGB space, the YUV domain can provide a more accurate representation of haze distribution characteristics and is better suited for image haze removal. Nevertheless, current YUV-based dehazing strategies often suffer from dull colors due to a lack of comprehensive consideration of color channels. To address this limitation, we propose an improved YUV domain-based image haze removal network (IHR-Net), which consists of a pre-trained model and a backbone network to enhance the Y channel and UV channels separately. Specifically, we utilize the pre-trained model to extract haze features that guide the backbone network in optimizing the UV channels. To increase the network’s attention to color-shifting areas, we introduce a haze attention module and integrate it into the backbone network. Experimental results on both synthetic and real-world daytime and nighttime images demonstrate the superior performance of IHR-Net.