Abstract The transition from 2 to 3D imaging marks a significant technological and conceptual leap across multiple domains, offering depth perception akin to real-world vision. A core enabler of this transition is depth computation through stereoscopic images, which is crucial for rendering realistic 3D scenes. One of the major challenges in this domain is the development of an effective stereoscopic image quality assessment (stereo-IQA) method that can evaluate the visual fidelity of stereo image pairs. This paper proposes a no-reference, goal-oriented stereo-IQA model designed to predict the effectiveness of a stereoscopic image pair in generating accurate depth maps. Unlike traditional IQA models that focus solely on perceptual distortions, our method estimates the suitability of stereo images for depth generation without the need for actual depth map computation. By using haar wavelet transforms, salient features are extracted from both left and right images and the model estimates the quality of the resulting depth by enabling a fast and efficient quality assessment tailored for depth-based applications. The model’s performance is evaluated using standard metrics: Pearson linear correlation coefficient (PLCC), spearman rank order correlation coefficient, Kendall rank order correlation coefficient, and root mean square error. Benchmark datasets including LIVE Phase-I, LIVE Phase-II, and the 3D22MX (augmented with four distortion types at five levels) were employed for experimental validation. The model achieves PLCC scores of 0.90 on LIVE Phase-I, 0.89 on Phase-II, and a remarkable 0.98 on 3D22MX, indicating strong correlations and low prediction errors across diverse distortion scenarios. These results affirm the robustness and practical utility of the proposed model in stereo-IQA for real-world 3D applications. Article highlights A novel stereo-IQA model that predicts the ability of stereoscopic image pairs to generate accurate depth maps, rather than focusing on perceptual distortions. Utilizes Haar wavelet transforms to extract salient features from stereo image pairs, enabling efficient depth quality prediction. Offers a fast and effective quality assessment method without the need for explicit depth map generation.
Kiruthika et al. (Mon,) studied this question.