Abstract
This paper introduces Saliency Guided Gradient Similarity (SGGS), a novel no-reference (NR) image blur metric designed for speed and perceptual accuracy. Unlike traditional NR metrics that lack a standardized output scale or fail to account for human visual focus, SGGS utilizes an intentional reblurring technique to create a pseudo-reference. By comparing the gradient similarity between the original and reblurred images—weighted by a saliency map—the method produces a score between 0 and 1 that correlates highly with human judgment. Experimental results show that SGGS outperforms existing methods in both accuracy and computational efficiency, processing high-resolution images in roughly one second.
Introduction
In Image Quality Assessment (IQA), "No-Reference" metrics are essential for evaluating images where the pristine original is unavailable. However, existing NR blur metrics face two primary challenges: a lack of a fixed scoring scale and a failure to model the Human Visual System (HVS). Humans do not perceive blur uniformly; we are far more sensitive to blur in salient regions (e.g., a subject's face) than in background areas. SGGS addresses these issues by providing an interpretable 0-1 scale and incorporating saliency-based pooling to prioritize perceptually important regions.
Methodology
The SGGS model operates through a three-step pipeline designed for maximum efficiency and accuracy:
- Intentional Reblurring: The algorithm applies a Gaussian low-pass filter to the input image. This creates a "pseudo-reference" version. The underlying logic is that reblurring a sharp image significantly alters its structure, whereas reblurring an already blurry image results in minimal change.
- Gradient Similarity: Rather than comparing raw pixels, the model calculates the similarity between the gradients of the input and the reblurred version. This focuses the assessment on edges and structural integrity, which are most affected by blur.
- Saliency-Based Pooling: To align with human perception, the model generates a Spectral Residual (SR) saliency map. This map identifies the most "interesting" parts of the image. The final blur score is a weighted average where salient regions contribute more heavily to the total score.
Experimental Results
The researchers evaluated SGGS against four gold-standard IQA databases: LIVE, CSIQ, TID2008, and TID2013. These datasets contain Gaussian-blurred images paired with subjective human scores (Mean Opinion Scores).
Accuracy and Consistency
SGGS demonstrated superior correlation with human judgment compared to established metrics like CPBD, S3, and MLV. On the TID2013 dataset, SGGS achieved a Pearson Linear Correlation Coefficient (PLCC) of 0.9267, significantly outperforming older reference methods. The gradient similarity approach ensures that the metric remains sensitive to structural changes across diverse image contents.
Computational Efficiency
Speed is the standout feature of SGGS. In benchmarks using high-resolution images (3264 × 2448), SGGS completed its assessment in just 1.05 seconds. In contrast, the S3 metric required nearly 460 seconds, and the MLV metric took 1.74 seconds. This efficiency makes SGGS suitable for real-time applications in mobile photography and automated image filtering.
Discussion and Conclusion
SGGS represents a significant step forward in NR-IQA by treating reblurring as a pseudo-reference and incorporating saliency to mimic human perception. While currently optimized for Gaussian blur, future work may extend the model to handle motion blur or chromatic aberration. By providing a standardized, fast, and perceptually accurate score, SGGS offers a robust solution for industrial-scale image quality monitoring.
Summary updated