Non-Uniform Illumination Underwater Image Enhancement via Minimum Weighted Error Entropy Loss

An effective enhanced unsupervised network based on minimum weighted error entropy (MWEE) loss is proposed for underwater image enhancement, which is one of the most challenging issues in computer vision. First, by the inspiration of Retinex theory, an underwater image is decomposed into non-uniform illumination and reflectance with shot noise. Then non-uniform illumination is modeled as an independent and piecewise identical (IPI) distribution, and shot noise in reflectance is seen as a single non-Gaussian distribution. Next, taking advantage of these two distributions, the MWEE criterion and its special case as training losses are embedded into a generative adversarial network (GAN) for piecewise uniformization of illumination and reflectance denoising. Experiments on underwater image enhancement datasets show the network enhanced by the proposed method obtains superior performance, and exhibits higher naturalness and better visual quality than several existing approaches.