Single Underwater Image Enhancement Based on LP-Norm Decomposition

Integration of ocean monitoring networks with artificial intelligence has become a popular topic for researchers. Artificial intelligence plays an important role in underwater image processing. For optical images captured in an underwater environment, the light scattering and absorption effect caused by the water medium results in poor visibility, such as blur and color casts. A novel approach is proposed herein to enhance the single underwater image with poor visibility. Similar to other image enhancement strategies built on fusion principles, our method also generates two input channels from the original degraded image, and these two channels are modulated by their corresponding weight measures. However, the main innovation of our method is that we propose a new multilevel decomposition approach based on l(p)-norm (p = 0, 1; 2) decomposition. According to the different sparse representation abilities of lp-norm to an image's spatial information, our approach decomposes the image into three levels: detail level, structure level, and illuminance level. Thus, these three levels can be manipulated separately. Because this new decomposition approach is based on image structural contents, rather than direct per-pixel downsampling that is utilized in traditional multi-resolution pyramid decomposition, it is more accurate and fiexible. Additionally, according to specific underwater imaging conditions, we carefully select two input channels and their three associated global contrast, local contrast, and saliency weight measures. Our method generates output with more accurate details and a better illuminant dynamic range. Generally, we are the first to impose an lp-norm-based decomposition strategy on underwater image restoration and enhancement. Extensive qualitative and quantitative evaluations demonstrate that our strategy yields better results than state-of-the-art algorithms.