Removal of line striping and shot noise from remote sensing imagery using a deep neural network with post-processing for improved restoration quality

Abstract

ABSTRACT Remote sensing imagery is generally prone to various radiometric distortions such as stripe noise, line loss, line or column drop out, banding, random bad pixels i.e. shot noise. These errors arrive due to on-board anomalies. These severely degrade the radiometric quality of the measured imagery and introduce a considerable level of incorrectness. Images with such a considerable level of radiometric incorrectness cannot be used directly for any image analysis application. It needs to be analyzed and pre-processed with Image Processing techniques before going to generate the data for the use of various applications. This paper presents a global residual learning method-based deep neural network (DNN) approach to automatically remove stripe noise of various types and shot noise as well, and hence improves the image quality of remote sensing imagery. The proposed method shows improved performance in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) quality measures when compared with the state-of-the-art methods. In addition, destriping by residual DNN is followed by a new post-processing step using multilevel wavelet decomposition and frequency domain filtering to remove the residual striping artefacts. A global residual learning method, batch normalization, mini batch selection and skip connection steps help speeding up the training process as well as boost the network performance.