Semantic Perceptual Image Compression

Abstract

This project implements a semantic, perceptual image compression pipeline that uses a convolutional model to identify regions of interest in an image and allocate compression quality accordingly. The result is intended to preserve visually important content while reducing overall image size.

The main entry point processes a single image from the command line, generates the compression output, and saves intermediate visualizations such as the ROI heatmap.

Directory Structure

• image_compressor.py: command-line entry point for compressing a single image
• compression.py: main compression pipeline and ROI-based quality allocation
• model.py: CNN and feature extraction logic used by the compressor
• util.py: image loading, resizing, normalization, and helper utilities
• jpeg_compression.py: baseline JPEG compression helper
• benchmark_test.py: benchmark script for comparing compression outputs
• get_metrics.py: metric collection and reporting utilities
• graph.py: plotting utilities for compression results
• Image_compression_deployment/: deployment-oriented copy of the project
• Image Quality Assessment Tools/: PSNR, SSIM, and related evaluation scripts

Installation

Install the required dependencies:

pip3 install -r requirements.txt

The project also expects CUDA-enabled TensorFlow dependencies if you are running the original GPU-based setup.

Commands

Run compression on a single image:

python3 image_compressor.py path/to/single_image.jpg

The script loads the input image, runs the compression pipeline, and writes output files to the project folders used by the code.

If you need the original invocation with elevated privileges, you can also run:

sudo python3 image_compressor.py path/to/single_image.jpg

Notes

• The compressor expects a file path to one image, not a directory or glob pattern.
• Output artifacts are written by the script into the local output/ directory and related generated files in the project root.