Wavelength-based Attributed Deep Neural Network for Underwater Image Restoration

Background: Underwater images, in general, suffer from low contrast and high color distortions due to the non-uniform attenuation of light as it propagates through water. In addition, degree varies with wavelength, resulting asymmetric traversing colors. Despite prolific works for underwater image restoration (UIR) using deep learning, above asymmetricity has not been addressed respective network engineering. Contributions: As first novelty, this article shows that attributing right receptive field size ( context ) based on range channel may lead a substantial performance gain task UIR. Further, is important suppress irrelevant multi-contextual features increase representational power model. Therefore, second we have incorporated an attentive skip mechanism adaptively refine learned features. The proposed framework, called Deep WaveNet , optimized traditional pixel-wise feature-based cost functions. An extensive set experiments carried out show efficacy scheme over existing best-published literature benchmark datasets. More importantly, demonstrated comprehensive validation enhanced images across various high-level vision tasks, e.g., semantic segmentation diver’s 2D pose estimation. A sample video exhibit our real-world available at https://tinyurl.com/yzcrup9n . Also, open-sourced framework https://github.com/pksvision/Deep-WaveNet-Underwater-Image-Restoration