Adherent Mist and Raindrop Removal from a Single Image Using Attentive Convolutional Network
This addresses a practical but underrated problem for vision systems like those in automotive or camera applications, though it appears incremental as it builds on existing attention mechanisms.
The paper tackles the problem of removing adherent mist and raindrops from single images, which degrades visibility in vision systems, and shows that the proposed method improves image quality both qualitatively and quantitatively while achieving state-of-the-art performance on related tasks.
Temperature difference-induced mist adhered to the glass, such as windshield, camera lens, is often inhomogeneous and obscure, easily obstructing the vision and severely degrading the image. Together with adherent raindrops, they bring considerable challenges to various vision systems but without enough attention. Recent methods for other similar problems typically use hand-crafted priors to generate spatial attention maps. In this work, we newly present a problem of image degradation caused by adherent mist and raindrops. An attentive convolutional network is adopted to visually remove the adherent mist and raindrop from a single image. A baseline architecture with general channel-wise attention, spatial attention, and multilevel feature fusion is used. Considering the variations and regional characteristics of adherent mist and raindrops, we apply interpolation-based pyramid-attention blocks to perceive spatial information at different scales. Experiments show that the proposed method can improve severely degraded images' visibility, both qualitatively and quantitatively. More applied experiments show that this underrated practical problem is critical to high-level vision scenes. Our method also achieves state-of-the-art performance on conventional dehazing and pure de-raindrop problems, in addition to our task of handling adherent mist and raindrops.