Image quality assessment by visual gradient similarity.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Image quality assessment (IQA) is crucial for evaluating digital images.
• Existing IQA models often lack accuracy or computational efficiency.
• Understanding human visual perception of image gradients is key to developing better IQA models.

Purpose of the Study:

• To introduce a novel full-reference IQA model based on multiscale visual gradient similarity (VGS).
• To incorporate human visual system properties related to image gradients into the IQA model.
• To evaluate the performance and efficiency of the VGS model against state-of-the-art metrics.

Main Methods:

• The VGS model employs a three-stage approach: global contrast registration, pointwise gradient similarity comparison, and multiscale pooling.
• Incorporated human visual system properties include Stevens' power law for gradient magnitude, quality uniformity, detection threshold, and frequency sensitivity.
• Model parameters were optimized using existing IQA databases and validated through cross-validation.

Main Results:

• The VGS model demonstrated high prediction precision and reliability.
• Experimental results confirmed VGS's competitive performance compared to current leading IQA metrics.
• The model exhibits simplicity and a low computational cost.

Conclusions:

• The multiscale VGS model provides an effective and efficient solution for full-reference image quality assessment.
• The incorporation of human visual system characteristics enhances the model's ability to predict subjective image quality.
• VGS presents a promising alternative for IQA applications requiring accuracy and speed.