Here I reproduce Pixel Error-Value Ordering with a CNN Predictor. The original is by Ankit Kumar Saini, Samayveer Singh, Signal, Image and Video Processing, vol. 19, 2025.
It trains a small neural predictor so the prediction errors shrink, which leaves more room for the payload at the same image quality.
I reimplemented the method from the paper and ran it from start to finish. Every number and every figure in the report is produced by the code in this repo, and reversibility is checked to be bit exact wherever the method claims it.
Everything runs with plain Python and a handful of common libraries. From this folder:
cd source_code
python3 run_experiment.py # runs the method, writes metrics and figures
python3 build_deliverables.py # rebuilds the IEEE report and the slidesYou need numpy, scipy, matplotlib, pillow, python-docx and python-pptx. Install them with pip install numpy scipy matplotlib pillow python-docx python-pptx. The report is exported to PDF with headless LibreOffice, so that has to be on the machine if you want the PDF rebuilt.
(the original paper stays on my machine and is not republished here, to respect its copyright)
ieee_report.docx/.pdf my IEEE format reproduction report
presentation.pptx a short summary deck
source_code/ the scripts that do the actual work
outputs/ metrics.json and the raw numbers behind the report
figures/ the plots and images the code produces
processing_notes.md what was reproduced, what was not, and the caveats
No results were made up. The only figures that come from the paper are the ones explicitly labelled as reported, kept so you can compare side by side. Where the exact image set or an unstated hyperparameter differs from the paper, the absolute figures can move a little, but the behaviour and the reversibility hold. The specifics for this paper are in processing_notes.md.