ALL Snow Removed: Single Image Desnowing Algorithm Using Hierarchical Dual-tree Complex Wavelet Representation and Contradict Channel Loss

Towards Efficient Single Image Dehazing and Desnowing

“…We also summarize the performance of our methods and other desnowing methods on three widely used benchmarks in Table 1. We compare the performance with previous state-of-the-art methods like HDCW-Net [7]. It's worth noting that (a) Input (b) DehazeNet [4] (c) FFA [30] (d) DA [35] (e)MSBDN [9] (f)PSD-FFA [8] (g)JSTASR [6] (h)HDCW [7] (i) DAN(Ours) (j) Ground-truth we also compare the performance of our approach with All-in-One network [22] which is trained to perform all the above tasks with a single model instance.…”

“…We compare the performance with previous state-of-the-art methods like HDCW-Net [7]. It's worth noting that (a) Input (b) DehazeNet [4] (c) FFA [30] (d) DA [35] (e)MSBDN [9] (f)PSD-FFA [8] (g)JSTASR [6] (h)HDCW [7] (i) DAN(Ours) (j) Ground-truth we also compare the performance of our approach with All-in-One network [22] which is trained to perform all the above tasks with a single model instance. Our method DAN-Net and its tiny version are also trained to perform both tasks using a single model instance.…”

“…Learning-based Desnowing Methods. For learning-based desnowing algorithms in the single image [6,7,10,22,26], the first presented desnowing network called DesnowNet [26] deals with the removal of translucent and opaque snow particles in sequence based on a multi-stage architecture. Li et al [22] develop an all-in-one network consisting of multiple task-specific encoders that can simultaneously handle multiple types of bad weather scenarios, such as rain, snow, haze and raindrops.…”

“…JSTASR [6] propose a novel snow formation model and size-aware snow network for single image desnowing which takes the veiling effect and variety of snow particles into consideration. HDCW-Net [7] performs snow removal by embedding the dual-tree wavelet transform into the architecture of the network and designing prior-based supervised loss called the contradict channel on the basis of the differences of snow and clean images. In the meantime, two large-scale snow datasets named snow removal in realistic scenario (SRRS) and comprehensive snow dataset (CSD) are respectively proposed in [6] and [7] for promoting the development of learning-based image desnowing methods.…”

“…We choose the widely used PSNR and SSIM as experimental metrcs to measure the performance of our networks. We train and test our expert networks on five large datasets : RESIDE [21], Haze4k [25], CSD [7], SRRS [6] and Snow100K [26], following the benchmarksetting of latest dehazing or desnowing methods [7,25] for authoritative evaluation. And we randomly choose 1,000 paired data from hazy or snowy datasets respectively to train our DAN-Net.…”