Abstract
In the hyperspectral image (HSI) denoising task, the real noise embedded in the HSI is always complex and diverse so that many model-based HSI denoising methods only perform well on some specific noisy HSIs. To enhance the noise adaptation capability of current methods, we first resort to the weighted HSI denoising model since its weight is capable of characterizing the noise in different positions of the image. However, the weight in these weighted models is always determined by an empirical updating formula, which does not fully utilize the noise information contained in noisy images and thus limits their performance improvement. In this work, we propose an automatic weighting scheme to alleviate this issue. Specifically, the weight in the weighted model is predicted by a hyper-weight network (i.e., HWnet), which can be learned in a bi-level optimization framework based on the data-driven methodology. The learned HWnet can be explicitly plugged into other weighted denoising models, and help adjust weights for different noisy HSIs and different weighted models. Extensive experiments verify that the proposed HWnet can help improve the generalization ability of a weighted model to adapt to more complex noise, and can also strengthen the weighted model by transferring the knowledge from another weighted model. Additionally, to explain the experimental results, we also theoretically prove the training error and generalization error upper bound of the proposed HWnet, which should be the first generalization error analysis in the low-level vision field as far as we know.
Abstract (translated)
URL
https://arxiv.org/abs/2301.06081
