[Paper Review] Image compression overview

Martin Prantl|arXiv (Cornell University)|Sep 14, 2014

Advanced Data Compression Techniques9 references5 citations

TL;DR

This paper provides a comprehensive overview of image compression techniques from the past decade, focusing on lossless and lossy methods. It evaluates prediction-based, wavelet-based, and neural network approaches, demonstrating that JPEG-LS and JPEG2000 achieve superior performance in lossless and lossy compression, respectively, with JPEG2000 outperforming others in PSNR at high compression ratios.

ABSTRACT

Compression plays a significant role in a data storage and a transmission. If we speak about a generall data compression, it has to be a lossless one. It means, we are able to recover the original data 1:1 from the compressed file. Multimedia data (images, video, sound...), are a special case. In this area, we can use something called a lossy compression. Our main goal is not to recover data 1:1, but only keep them visually similar. This article is about an image compression, so we will be interested only in image compression. For a human eye, it is not a huge difference, if we recover RGB color with values [150,140,138] instead of original [151,140,137]. The magnitude of a difference determines the loss rate of the compression. The bigger difference usually means a smaller file, but also worse image quality and noticable differences from the original image. We want to cover compression techniques mainly from the last decade. Many of them are variations of existing ones, only some of them uses new principes.

Motivation & Objective

• To analyze and compare recent advancements in image compression techniques over the last decade.
• To evaluate the performance of lossless methods such as PNG, CALIC, JPEG-LS, and JPEG2000 using standard image sets.
• To assess lossy compression techniques including JPEG2000, JPEG, quadtree-based, and neural network-based methods in terms of PSNR and compression ratio.
• To investigate the impact of image content (e.g., edges, smooth regions) on compression efficiency and quality.
• To highlight the limitations and challenges in adopting newer compression methods despite their theoretical advantages.

Proposed method

• Uses a comparative evaluation framework on standard 512x512 and 256x256 test images (Lenna, Trui, Airplane, Crowd, Landsat).
• Employs PSNR and bits per pixel (bpp) as primary metrics for lossy compression evaluation.
• Applies lossless compression using dictionary-based (PNG, GIF), prediction-based (CALIC, JPEG-LS), and transform-based (JPEG2000) methods.
• Evaluates prediction-based methods using gradient adaptive prediction (GAP), median edge detection (MED), and neural network predictors.
• Tests neural network-based compression using feed-forward networks trained via back-propagation on pixel neighborhoods.
• Analyzes compression efficiency through histogram similarity and image content characteristics (e.g., edge density, smoothness).

Experimental results

Research questions

• RQ1 How do modern lossless image compression methods compare in terms of compression ratio and speed?
• RQ2 What is the impact of image content (e.g., edges, texture, uniform regions) on the performance of different compression algorithms?
• RQ3 How does JPEG2000 compare to other lossy compression methods in terms of PSNR and compression efficiency across varying bitrates?
• RQ4 Why do some neural network-based compression methods show inconsistent performance at low bitrates?
• RQ5 To what extent do near-lossless methods like JPEG-LS simulate lossy compression while maintaining high fidelity?

Key findings

• JPEG2000 achieved the best PSNR performance in lossy compression, significantly outperforming JPEG and other methods, especially at high compression ratios.
• For lossless compression, JPEG-LS and the DWT-based method from [18] delivered the best results, with JPEG-LS outperforming PNG and lossless JPEG2000 in both speed and compression ratio.
• The MED predictor in JPEG-LS effectively reduces prediction errors, especially along strong edges, improving compression efficiency.
• Neural network-based compression showed inconsistent results at low bitrates, likely due to training instability, despite promising performance at higher bitrates.
• Images with similar histograms (e.g., Airplane and Crowd) showed comparable compression ratios, indicating strong correlation between image statistics and compression efficiency.
• Contour-based compression [37] performed poorly on natural images like Lenna and Trui, confirming its unsuitability for general-purpose image compression.