Opportunity
High Dynamic Range (HDR) images offer realistic visual experiences with broader luminance and richer structural details than standard LDR images. However, existing Image Quality Assessment (IQA) methods designed for LDR images are not directly applicable to HDR images due to fundamental differences in data distribution and luminance range. As HDR content becomes increasingly common in TV, photography, and streaming, there is a pressing need for objective IQA models specifically tailored for HDR images. Such models must align with the Human Visual System (HVS), which has non-linear sensitivity to spatial frequency—responding strongly to certain mid-range frequencies while being less sensitive to very low or very high frequencies. No existing IQA method effectively captures both local texture details and global frequency characteristics in a way that matches HVS perception for HDR content.
Technology
This patent presents a Local and Global Frequency feature-based Model (LGFM) for full-reference HDR image quality assessment. The system takes a pair of reference and distorted HDR images. First, an image pre-processing module converts both images to a perceptual space (using Perceptual Unit/PU coding) to map wide HDR luminance to HVS perceptual range.
A feature extraction module then extracts two complementary frequency features. For local frequency features, an odd log-Gabor filter extracts horizontal and vertical edge information, capturing texture and structure details. A spatial mask gives higher weights to high-luminance regions where HDR provides superior detail. For global frequency features, a bandpass Butterworth filter simulates the Contrast Sensitivity Function (CSF) of the HVS, extracting a frequency map and phase map from the image spectrum after Discrete Fourier Transform (DFT).
A comparison module generates similarity maps for both local and global features between the reference and distorted images. Finally, a scoring module performs weighted feature pooling to calculate local and global similarity scores, multiplying them to produce a final IQA score. Extensive experiments on four benchmark datasets demonstrate that LGFM outperforms state-of-the-art HDR IQA methods (HDR-VDP-3, HDR-VQM) and LDR methods adapted for HDR.
Advantages
- Tailored for HDR: Specifically designed for HDR image characteristics (wide luminance, rich texture), unlike LDR methods that perform poorly on HDR.
- HVS-Aligned Local Features: Odd log-Gabor filter captures edge/structure information that closely matches HVS response.
- Luminance-Weighted Focus: Spatial mask prioritizes high-luminance regions where HDR excels, improving perceptual relevance.
- CSF-Simulated Global Features: Bandpass Butterworth filter models the HVS's frequency sensitivity curve, capturing perceptually important frequency intervals.
- Outperforms Existing Methods: Achieves top SROCC, KROCC, and RMSE scores across multiple HDR IQA benchmarks, outperforming HDR-VDP-3 and HDR-VQM.
- Computationally Efficient: Requires fewer parameters and less running time than competing methods.
Applications
- HDR Image/Video Compression Optimization: Evaluating and tuning codec parameters to maximize compression ratio while preserving perceptual quality.
- HDR Content Generation: Assessing the quality of tone-mapped or reconstructed HDR images from LDR sources.
- Broadcast & Streaming: Monitoring HDR video quality in real-time transmission to ensure viewer experience.
- Display Calibration & Testing: Quantifying how different HDR displays reproduce reference image quality.
- Photography & Post-Production: Providing objective quality feedback during HDR image editing, fusion, or retouching.
