Learned Image Codec Optimized for Speed and Perceptual Quality
A recently developed learned image codec demonstrates a notable enhancement in balancing speed with perceptual quality. This research thoroughly examines essential modeling decisions for effective learned image codecs that are optimized for both perceptual quality and runtime, incorporating innovative methods. By performing a performance-driven neural architecture search across millions of backbone configurations, the study identifies models that achieve on-device runtime objectives while optimizing compression performance according to perceptual metrics.
Key facts
- Learned codecs can be optimized for human visual system
- Perceptual yet practical image codec not yet proposed
- Comprehensive study of modeling choices for practical learned image codec
- Joint optimization for perceptual quality and runtime
- Includes several novel techniques in ablations
- Performance-aware neural architecture search over millions of backbone configurations
- Target on-device runtime while maximizing compression performance
- New codec achieves improved speed-perceptual quality tradeoff
Entities
—