Grunnleggende konsepter
Learned Image Compression with ROI-Weighted Distortion and Bit Allocation enhances perceptual quality through region-based bit allocation.
Sammendrag
Abstract:
TLIC method for image compression focuses on perceptual quality enhancement.
Adversarial loss used to generate realistic textures.
Region of Interest (ROI) mask guides bit allocation.
Introduction:
Learned image compression surpasses non-learned codecs.
Perceptual quality improvement using GANs and learned metrics.
TLIC proposed for better compression focusing on authenticity over vividness.
Method:
Overview:
TLIC architecture similar to Ma et al. with gain units for rate control.
ROI-Weighted Distortion and Bit Allocation:
Saliency map used as ROI map for better allocation.
RMformer generates smoothed ROI maps for bit-allocation guidance.
Adversial Training:
Adversarial loss employed for realistic texture generation at low bit-rate.
Variable Rate Adaptation:
Gain units adjust quantization step for continuous rate adaptation.
Entropy Model:
Simplified entropy model with local context modules employed.
Training:
320x320 patches used during training with a batch size of 8.
Conclusion:
TLIC method employs ROI-weighted rate allocation and distortion.
U-net based discriminator enhances feedback in adversarial optimization.
Gain units and inverse gain units used for target bit-rate achievement.