Wavelet-Supervision Convolutional Neural Network for Restoration of JPEG-LS Near Lossless Compression Image

Abstract

JPEG-LS near lossless compression algorithm is widely used in remote sensing image compression. However, the run-length coding of the algorithm results in horizontal stripe distortion of the decompressed image, which will greatly affect the quality of remote sensing image. In order to solve such distortion and restore the image, a wavelet-supervision convolutional neural network (WSCNN) with large receptive field is proposed. WSCNN can make full use of the information both in spatial and frequency domains. With translation invariance, convolutional neural network (CNN) is very adept at extracting features in pixel space. To further explore spatial information, we enlarge the receptive field of our WSCNN. Alternatively, wavelet coefficients show promising of frequency information digging, we adopt them to supervise our WSCNN. With this wavelet-supervision, WSCNN can focus on the horizontal stripe distortion in frequency domain. Besides, we have collected a dataset, it consists of original remote sensing images at a resolution of $4096\times 4096$ and corresponding JPEG-LS near-lossless compression images as data pairs. Subjective and objective experiments have verified the effectiveness of our WSCNN for the restoration of LPEG-LS near-lossless compression images.