QFCOI - Enhancing Objective Quality Assessment for Compressed Omnidirectional Images with Fusion of Measures

Abstract

This paper introduces Quality Fusion of Compressed Omnidirectional Images (QFCOI), an enhanced objective quality assessment method for 360° images. QFCOI integrates linear fusion of feature metrics. Established conventional state-of-the-art measures were analyzed to select specific ones for effective fusion and to mitigate eventual overfitting. The feature metrics selection was based on statistical performance and ability to capture key aspects of image quality, including structural preservation, visual information fidelity, and artifact sensitivity. To optimize predictive performance of QFCOI, a genetic algorithm was utilized to determine optimal weight coefficients, maximizing the monotonic correlation with subjective quality scores. The QFCOI performance was validated through correlation coefficients, statistical significance testing, and Receiver Operating Characteristic (ROC) analyses. Results confirmed that QFCOI outperformed other conventional metrics, achieving the highest performance on the OMNIQAD dataset across multiple emerging compression algorithms, including High Efficiency Image File Format (HEIC), Joint Photographic Experts Group XL (JPEG XL), and AV1 Image File Format (AVIF). Further, validation on the relevant public CVIQ dataset for HEIC-compressed images confirmed the robustness of QFCOI, and the method modification on only 25 % of the data, achieved the best performance on the CVIQ dataset. These results highlight the generalization properties and versatility of QFCOI. In contrast to learning-based models, the proposed method offers a transparent and interpretable alternative while achieving high accuracy and statistical reliability in objective quality assessment.This paper introduces Quality Fusion of Compressed Omnidirectional Images (QFCOI), an enhanced objective quality assessment method for 360° images. QFCOI integrates linear fusion of feature metrics. Established conventional state-of-the-art measures were analyzed to select specific ones for effective fusion and to mitigate eventual overfitting. The feature metrics selection was based on statistical performance and ability to capture key aspects of image quality, including structural preservation, visual information fidelity, and artifact sensitivity. To optimize predictive performance of QFCOI, a genetic algorithm was utilized to determine optimal weight coefficients, maximizing the monotonic correlation with subjective quality scores. The QFCOI performance was validated through correlation coefficients, statistical significance testing, and Receiver Operating Characteristic (ROC) analyses. Results confirmed that QFCOI outperformed other conventional metrics, achieving the highest performance on the OMNIQAD dataset across multiple emerging compression algorithms, including High Efficiency Image File Format (HEIC), Joint Photographic Experts Group XL (JPEG XL), and AV1 Image File Format (AVIF). Further, validation on the relevant public CVIQ dataset for HEIC-compressed images confirmed the robustness of QFCOI, and the method modification on only 25 % of the data, achieved the best performance on the CVIQ dataset. These results highlight the generalization properties and versatility of QFCOI. In contrast to learning-based models, the proposed method offers a transparent and interpretable alternative while achieving high accuracy and statistical reliability in objective quality assessment.