PreSPA: New IQA method rivals full-reference accuracy with minimal reference data
Only one scalar from reference image needed – PreSPA achieves near full-reference quality assessment.
PreSPA (Partial-Reference Structural Prediction Approach) offers a new way to assess image quality without needing the full reference. It separates quality into two complementary indices: a structure-aware index that captures degradation via Hermite-Gauss predictions of the distorted gradient field and its curvature angular variance, and a texture-sensitive index that estimates local noise from energy differences on strong-edge regions. Crucially, only a single scalar μ is extracted from the reference image per pair, minimizing the reference footprint to its information-theoretic limit. The final score is computed via an affine fusion with just three interpretable parameters, making the method compact and transparent.
Evaluated on six standard IQA benchmarks, PreSPA consistently rivals or exceeds leading no-reference approaches, and in several cases matches the accuracy of full-reference models. The method embeds viewing distance into the operator scale and requires no dataset-specific calibration, offering a practical balance between minimal reference data and high fidelity. This could enable efficient quality monitoring in streaming, compression, and real-time video applications where full reference is unavailable or costly.
- Only a single scalar μ from the reference is needed, reducing reference footprint to a minimum.
- Uses Hermite-Gauss structural prediction and texture deviation for dual-indices quality assessment.
- Rivals top no-reference methods and matches full-reference accuracy on 6 standard benchmarks.
Why It Matters
Enables high-quality image assessment with minimal reference data, ideal for streaming and real-time applications.