PSNR vs Subjective JPEG Quality Correlation

Peak Signal-to-Noise Ratio (PSNR) is one of the most widely used mathematical metrics for evaluating image reconstruction quality, yet its correlation with subjective human perception in JPEG compression is inconsistent. While a higher PSNR generally suggests a cleaner image, it measures absolute pixel error rather than visual significance. This article explores how PSNR aligns with human visual perception, where the metric fails to reflect subjective JPEG quality, and why structural distortions often evade simple numerical evaluation.

The Mechanics of PSNR in JPEG Compression

PSNR measures the ratio between the maximum possible power of an image signal and the power of corrupting noise, expressed on a logarithmic decibel (dB) scale. It relies directly on Mean Squared Error (MSE), which calculates the squared differences between corresponding pixels in the original and compressed JPEG images:

\[\text{MSE} = \frac{1}{MN} \sum_{i=0}^{M-1} \sum_{j=0}^{N-1} [I(i,j) - K(i,j)]^2\]

\[\text{PSNR} = 10 \cdot \log_{10} \left( \frac{MAX_I^2}{\text{MSE}} \right)\]

Because JPEG compression utilizes the Discrete Cosine Transform (DCT) and quantizes frequency coefficients, error is not distributed uniformly across the image. PSNR aggregates these errors into a single scalar value without considering spatial arrangement or visual relevance.

Where PSNR Correlates Well with Subjective Quality

PSNR correlates reasonably well with subjective evaluations—such as Mean Opinion Scores (MOS)—under tightly controlled conditions:

• Intra-Image Comparisons: When evaluating the same image across varying JPEG quality factors (e.g., Q10 vs. Q50 vs. Q90), PSNR tracks monotonically with perceived quality. As the quality factor increases, MSE decreases, PSNR rises, and human observers agree the image looks better.
• Extreme Compression Ranges: At very low compression ratios (PSNR > 38 dB), images appear nearly identical to the human eye. At severe compression ratios (PSNR < 25 dB), artifacts are universally objectionable. In these extreme bands, PSNR reliably reflects perceived fidelity.

Where the Correlation Breaks Down

The correlation between PSNR and subjective quality weakens significantly when comparing different images or evaluating intermediate JPEG compression artifacts.

1. Inability to Penalize Structured Artifacts

JPEG compression divides images into 8x8-pixel blocks. At moderate-to-low quality settings, this produces distinct block boundary discontinuities (blocking artifacts) and ringing around sharp edges. The Human Visual System (HVS) is exceptionally sensitive to regular, structural patterns like grid lines. PSNR treats a 5-unit pixel shift along an 8x8 block boundary the same as a 5-unit shift scattered randomly across textured grass, whereas a human observer will notice and penalize the block boundary immediately.

2. Content Dependency and Visual Masking

Human perception relies heavily on contrast masking and visual context:

• High-Texture Images: Noise in busy areas (foliage, fabric, gravel) is masked by the existing texture. An image can suffer high MSE and a low PSNR in textured regions while remaining visually acceptable to a viewer.
• Low-Frequency/Smooth Images: In flat areas like blue skies or skin tones, even minor luminance errors or color banding are immediately visible. An image can achieve a high PSNR despite containing noticeable false contours.

Consequently, two distinct JPEG images with an identical PSNR of 32 dB can yield drastically different subjective MOS ratings; one may appear visually clean, while the other exhibits glaring artifacts.

Practical Implications

While PSNR provides a fast, mathematically trackable benchmark for optimization algorithms, it should not be used in isolation to judge final visual output for human consumption. Metrics designed with human perceptual models—such as the Structural Similarity Index (SSIM), Multi-Scale SSIM (MS-SSIM), or learned perceptual metrics like LPIPS—correlate far more reliably with subjective human ratings by weighting edge degradation, contrast changes, and localized luminance masking.