A Perceptually Lossless Image Compression Scheme Based on JND Refinement by Neural Network
Wen‐Nung Lie, Wen-Ching Liu · 2010
In this paper, we propose a JND (Just Notice Distortion)-loss less image compression scheme that can improve the compression performance for JPEG-LS, while maintaining the image perceptual quality simultaneously. JND-loss less can be easily achieved by setting the quantization step size (QSS) to be double the JND value. However, dynamic JNDs make the coding of varying QSSs difficult and the JND estimated at decoder is often inaccurate due to incomplete background information therein. Here, we propose to use a neural classifier at both the encoder and decoder to correct the JND mismatch and make the coding of varying QSSs unnecessary. Experiments show that our proposed JND-refinement scheme is capable of increasing the compression performance by up to 15% (w.r.t. JPEG-LS) and the proposed neural classifier is capable of correcting 94% of JND mismatch.