Image Compression Quality Measurement: Correspondence Between PSNR, SSIM and Quality Parameters

PSNR, short for Peak Signal-to-Noise Ratio, is mathematically defined as 10 multiplied by the base-10 logarithm of the reciprocal of the mean squared error between the original image and the compressed image divided by the peak signal power. It is calculated based on the global error of pixel values, and its unit is dB.

A higher PSNR value indicates smaller pixel error, but this metric does not match human visual perception. A compressed image with high PSNR may have visible blocking artifacts, while the difference in an image with low PSNR is hard for the human eye to detect. This is its core limitation.

SSIM, short for Structural Similarity Index Measure, ranges from 0 to 1. A value closer to 1 means smaller structural difference between the compressed image and the original image. This metric compares images from three dimensions: luminance, contrast and structure, instead of only counting pixel errors.

The human eye is more sensitive to structural information of an image than random pixel errors, and SSIM assigns higher weight to structural distortion, so its score is closer to the actual perception result of the human eye.

The quality parameter range of JPG, WebP and AVIF is 1 to 100 for all three formats, and all have non-linear characteristics. When the quality parameter increases from 90 to 100, the file size increases significantly, but the quality improvement is very limited.

When the quality parameter is below 60, every 10-unit increase in the parameter brings a far larger improvement in image quality than an increase from 80 to 90. Parameter changes in the low-quality range affect image quality more significantly, while the marginal benefit decreases rapidly in the high-quality range.

This test selected 10 conventional color images from different scenarios, covering landscapes, portraits, and mixed text-image content, and tested the SSIM values of the three formats with different quality parameters respectively.

For all three formats, SSIM values in the quality parameter range 75 to 85 are between 0.98 and 0.995. Most observers cannot distinguish the difference between compressed images in this range and the original images, and the file size is 40% to 60% smaller than at quality 100.

Chroma subsampling 4:2:0 is the default chroma sampling method for most lossy compression formats. This scheme halves the resolution of chroma channels, reducing the amount of chroma data by approximately 50%.

For screen shots that contain small, clear text, 4:2:0 sampling can cause blurring artifacts on the edges of colored text. If you need to keep text clear, you can choose a compression format that uses 4:4:4 sampling, such as PNG, or WebP with 4:4:4 enabled.