Abstract
Omnidirectional image (ODI) super-resolution is critical for immersive Virtual Reality (VR) experiences. However, standard convolutional neural networks (CNNs) struggle with the geometric distortions inherent in Equirectangular Projections (ERP). This paper introduces ODA-SRN (Omnidirectional Distortion-Adaptive Super-Resolution Network), which incorporates a "Dual Enhancement" strategy. By adapting convolution kernels and upsampling windows to spherical geometry, ODA-SRN achieves state-of-the-art performance with significantly higher efficiency than previous methods.
Introduction
The demand for high-resolution 360-degree content in VR applications is growing rapidly. Most ODIs are stored using ERP, which maps a spherical surface onto a rectangular plane. This process introduces non-uniform distortion: pixels near the poles are significantly stretched compared to those at the equator. Standard CNNs, designed for planar images, apply uniform filters across the entire image, leading to blurred details and artifacts in highly distorted regions.
Methodology: Dual Enhancement Strategy
ODA-SRN addresses distortion through two primary architectural innovations:
1. Distortion Attention Block (DAB)
Instead of static convolution, the DAB employs multi-segment parameterized convolution. The image is partitioned into horizontal segments based on latitude. A lightweight routing network generates dynamic, custom filters for each segment, allowing the model to specialize its "experts" for different levels of distortion (e.g., stretched poles vs. stable center).
2. Distortion Offset Upsampling (DOU) Layer
Standard upsampling methods like pixel shuffling do not account for the varying physical area represented by pixels at different latitudes. The DOU layer utilizes a Jacobian matrix to model the sphere-to-plane mapping mathematically. It introduces a "sampling window correction" that dynamically adjusts the receptive field, pulling features from less-distorted neighboring regions to compensate for the sparse information at the poles.
3. Distortion-Aware Loss
The training process is guided by a loss function weighted by a "zoom factor" ($\delta$). This ensures the network prioritizes reconstruction accuracy in regions where projection distortion is most severe, leading to better overall spherical quality.
Experimental Results
The model was evaluated on a curated dataset of 2,600 high-quality ODIs. Performance was measured using standard PSNR/SSIM and spherical-specific metrics like Weighted-to-Spherical PSNR (WS-PSNR).
- State-of-the-Art Performance: On $\times4$ upscaling, ODA-SRN achieved a WS-PSNR of 29.47 dB, outperforming the Transformer-based OSRT.
- Efficiency: ODA-SRN uses only 8.9M parameters—a 92% reduction compared to OSRT's 119.3M.
- Speed: The model processes images in approximately 0.298 seconds, making it nearly 7x faster than OSRT and 11x faster than EDSR.
Conclusion
ODA-SRN demonstrates that incorporating geometric priors into network design is more effective than simply increasing model depth for ODI super-resolution. By adapting both the feature extraction (convolution) and reconstruction (upsampling) stages to projection distortion, the model provides a high-fidelity, computationally efficient solution suitable for real-world VR deployment. Future work will explore extending these principles to video super-resolution.
Summary updated