A Multiscale Approach to Deep Blind Image Quality Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Quality Assessment

Background:

• Accurate perceptual quality measurement is vital for multimedia applications.
• Full-reference image quality assessment (FR-IQA) uses reference images for high performance.
• No-reference image quality assessment (NR-IQA), or blind image quality assessment (BIQA), lacks reference images, posing a significant challenge.

Purpose of the Study:

• To develop an advanced NR-IQA method that overcomes limitations of spatial-domain focused approaches.
• To enhance image quality assessment by incorporating frequency domain information.
• To propose a novel multiscale deep blind image quality assessment (BIQA, M.D.) method.

Main Methods:

• The proposed method, BIQA, M.D., employs spatial optimal-scale filtering analysis.
• Images are decomposed into multiple spatial frequency bands using multiscale filtering, inspired by human visual system's multi-channel behavior and contrast sensitivity.
• Features extracted from these bands are mapped to subjective quality scores using a convolutional neural network.

Main Results:

• Experimental results demonstrate that the BIQA, M.D. method achieves competitive performance compared to existing NR-IQA techniques.
• The method shows strong generalization capabilities across different image datasets.
• The approach effectively utilizes information from various frequency bands, improving prediction accuracy.

Conclusions:

• The developed multiscale deep blind image quality assessment method offers a promising solution for NR-IQA.
• Integrating spatial and frequency domain analysis leads to more robust and accurate image quality prediction.
• The BIQA, M.D. method provides a valuable tool for applications requiring reliable perceptual quality assessment without reference images.