Trainable spectral difference learning with spatial starting for hyperspectral image denoising

Because of the limited reflected energy and incoming illumination in an individual band, the reflected energy captured by a hyperspectral sensor might be low and there is inevitable noise that significantly decreases the performance of the subsequent analysis. Denoising is therefore of first importance in hyperspectral image (HSI) analysis and interpretation. However, most HSI denoising methods remove noise with the important spectral information being severely distorted. This paper presents an HSI denoising method using trainable spectral difference learning with spatial initialization (called HDnTSDL) aimed at preserving the spectral information. In the proposed HDnTSDL model, a key band is automatically selected and denoised. The denoised key band acts as a starting point to reconstruct the rest of the non-key bands. Meanwhile, a deep convolutional neural network (CNN) with trainable non-linearity functions is proposed to learn the spectral difference mapping. Then, the rest of the non-key bands are denoised under the guidance of the learned spectral difference with the key band as a starting point. Experiments have been conducted on five databases with both indoor and outdoor scenes. Comparative analyses validate that the proposed method: (i) presents superior performance in spatial recovery and spectral preservation, and (ii) requires less computational time than state-of-the-art methods.