Bayesian Approach in a Learning-Based Hyperspectral Image Denoising Framework

Hyperspectral images are corrupted by a combination of Gaussian-impulse noise. On one hand, the traditional approach handling denoising problem using maximum posteriori criterion is often restricted time-consuming iterative optimization process and design hand-crafted priors to obtain an optimal result. other discriminative learning-based approaches offer fast inference speed over trained model; but highly sensitive noise level used for training. A model with loss function which does not accord Bayesian degradation leads sub-optimal results. In this paper, we training paradigm emphasizing role functions; similar as observed in model-based methods. As result; functions derived setting employed neural network boosts performance. Extensive analysis experimental results on synthetically real hyperspectral dataset suggest potential applicability proposed technique under wide range homogeneous heterogeneous noisy settings.