Official PyTorch implementation of the paper "LWANet: Lightweight Wavelet Attention Network for 3D-HEVC Compressed Depth Map Enhancement".
Depth maps play a crucial role in synthesized view generation for 3D Video (3DV) systems. However, 3D-HEVC compression inevitably introduces artifacts like blockiness and ringing, which degrade the quality of synthesized views and cause visual discomfort in VR/AR applications.
LWANet is an ultra-lightweight deep learning solution that:
- Employs a dual-stream architecture to capture global geometric structures and local details.
- Integrates a wavelet attention mechanism to reduce spatial redundancy while preserving high-frequency edges.
- Achieves state-of-the-art performance with only 461K parameters (a 60% reduction compared to existing methods).
- Specific Design for Depth Maps: Unlike general image enhancement, LWANet focuses on the geometric characteristics of depth maps to ensure high-fidelity virtual view synthesis.
- Lightweight Feature Extraction (LFE): Utilizes Wavelet Channel Aggregation (WCA) and Wavelet Residual Blocks (WRB) for efficient multi-scale analysis.
- Residual Feature Enhancement (RFE): A hierarchical branch that progressively refines features using Multi-level Wavelet Residual Blocks (MWRB).
- Efficiency-Performance Balance: Optimized for real-time multimedia systems with minimal computational overhead.
The pipeline consists of three main stages:
- LFE Module: Extracts multi-scale features via wavelet decomposition and adaptive channel weighting.
- RFE Module: Progressively enhances the extracted features to recover lost edge details.
- Reconstruction Stage: Aggregates all refined information to generate the final high-quality depth map.
Important
The full source code and pre-trained models will be released soon upon the official acceptance of the paper. ---
If you use this code or model in your research, please cite our paper:
The full source code and pre-trained models will be released soon upon the official acceptance of the paper. ---