Remote Sensing Image Mixed Noise Denoising with Noise Parameter Estimation

Abstract

Images acquired by optical remote sensing often suffer from noise, due to low-light conditions and long distances from the object. Noise introduces undesirable effects such as artifacts, unrealistic edges, unseen lines and corners, blurred objects and disturbed background scenes, and decreasing the performance of visual and computational analysis.

The great majority of denoising methods only focus on single noise type, and mostly assume additive Gaussian white noise. However, in practice, according to the different sources of noise during the capturing process, the real image noise may be composed of noises from several different distributions, i.e., Poisson distribution from shot noise and dark current, Gaussian distribution from readout noise, etc. Thus, it is hard to model complex noise with a single noise model, which therefore limits the practical performance of image denoiser for real captured images. To address this problem, we propose a deep residual network with parameter estimation to model and remove mixed noises of remote sensing data, with no noise type and noise level information.

In this chapter, a deep convolutional neural network based framework is proposed to implement both noise estimation and noise removing. The framework is composed of noise estimator module and denoiser module. The noise estimator module is used to estimate the noise distribution characteristics of the input image, which are used to synthesize the noisy dataset for the training of the denoiser module. After training, the denoiser module can then denoise the input image specifically for that noise characteristic of the input image. Besides, we use some advanced designment to improve the performance of the network like residual learning, skip connections, and perceptual loss. Our model works for both given noise types and others without taking into account in training process, and for the latter case we test the model on the noise mixed with different types, i.e., shot noise, Gaussian noise, salt and pepper noise, speckle noise, dark current, and quantization noise, covering most of the common noise types in practical situation. The estimated noise distribution according to the proposed estimator is compared with the ground truth to validate the performance of the estimator. We also compare the proposed method with the existing denoising methods, including both the traditional algorithm (i.e., BM3D (Dabov et al., IEEE Trans Image Process 16:2080–2095, 2007)) and the deep neural network (i.e., DnCNN (Zhang et al., IEEE Trans Image Process 26:3142–3155, 2017)). Experimental results showed that our method outperforms these state-of-the-art methods in both PSNR (peak signal to noise ratio) and SSIM (structural similarity index).