Paper Available at: https://arxiv.org/abs/2102.00502
Estimate the quantization error in lossy compression and find an optimal inverse kernel using linear regression rather than original IDCT kernel.
Our learnt optimal inverse kernels have a low dependency on training images. The kernel trained at certain quality factor N can be used to decode images encoded by a quality factor near to it. It can achieve 0.1-0.2dB PSNR increment on DIV2K dataset @N=50 and more than 0.3dB PSNR increment on Normalized Brodatz textures from MBT dataset @N=90.
More details can be found in our paper.
Use ./src/main.py to train the optimal inverse kernel on training images. Then copy the kernel value in ./kernel/ folder to inv_K in ./src/jpeg-6b-our/jidctflt.c.
Use ./src/run.sh to encode and decode all provided images.
./src/plot.py is used to run the evaluation result and plot the figures used in this paper.