WaveDiff: Underwater Visual Data Enhancement with Wavelet-Accelerated Diffusion

Unmanned underwater vehicles generally rely on high quality visual data and low latency for the applications of monitoring, exploration and search. Several methods have been proposed to improve underwater visibility by enhancing images taken by cameras using generative methods such as diffusion models. Although diffusion models have proven very useful for visual data enhancement, they enhance images iteratively, making this process computationally inefficient. Current methods primarily focus on improving visibility without addressing operation latency or speed, rendering these methods unsuitable for low bandwidth or fast information tracking. To overcome this, we propose using discrete wavelet transform (DWT) to decompose the image into four frequency components, reducing the spatial dimension. Out of these four components, we use only one as part of the diffusion process. At the end of the diffusion process, we apply the inverse discrete wavelet transform to obtain the enhanced image. This simple strategy helps reduce processing time by 50% with minimal loss in Peak Signal to Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) on the LSUI dataset.