Belangrijkste concepten
A novel neural image compression architecture that utilizes auxiliary information to predict multi-scale features of the original image, enabling efficient encoding of only the feature residuals in the main network.
Samenvatting
The proposed method introduces a new neural image compression architecture that consists of an auxiliary coarse network and a main network. The auxiliary coarse network encodes auxiliary information and predicts multi-scale features as an approximation of the original image. The main network then encodes only the residual between the predicted features and the original image features, leading to efficient compression.
Key highlights:
- The auxiliary coarse network predicts multi-scale features of the original image using the auxiliary information, and the main network encodes the residual between the predicted features and the original image features.
- The Auxiliary info-guided Feature Prediction (AFP) module effectively predicts the original image features using global correlation.
- The Context Junction module refines the predicted auxiliary features and implicitly subtracts them from the original image features.
- The Auxiliary info-guided Parameter Estimation (APE) module predicts the approximation of the latent vectors and estimates their probability distribution using the auxiliary information.
- Extensive experiments demonstrate that the proposed model outperforms state-of-the-art neural image compression methods, achieving up to 19.49% higher rate-distortion performance on the Tecnick dataset compared to the VVC codec.
Statistieken
The proposed model achieves a 19.49% higher rate-distortion performance than VVC on the Tecnick dataset.
The proposed model saves an average of 19.49% bits for the same PSNR quality on the Tecnick dataset compared to VVC.
Citaten
"Inspired by neural video compression structures, we introduce a new prediction architecture for neural image compression."
"To further leverage our new structure, we propose Auxiliary info-guided Feature Prediction (AFP) module that uses global correlation to predict more accurate predicted features."
"Finally, we introduce Auxiliary info-guided Parameter Estimation (APE) module, which predicts the approximation of the latent vector and estimates the probability distribution of these residuals."