Context-adaptive neural network based prediction for image compression.

Area of Science:

• Computer Vision
• Machine Learning
• Image Compression

Background:

• Intra image prediction is crucial for video compression efficiency.
• Existing methods like H.265 intra prediction modes have limitations in modeling complex textures.
• Neural networks offer potential for improved intra prediction.

Purpose of the Study:

• To develop a novel set of neural network architectures (PNNS) for intra image prediction.
• To adaptively select between fully-connected and convolutional neural networks based on image block size.
• To enhance the performance of video codecs through improved intra prediction.

Main Methods:

• Designed Prediction Neural Networks Set (PNNS) using fully-connected and convolutional neural networks.
• Implemented adaptive selection of neural network types based on image block size.
• Utilized masks of random sizes during training for context adaptability.
• Integrated PNNS into the H.265 video compression codec.

Main Results:

• Achieved PSNR-rate performance gains ranging from 1.46% to 5.20% when integrating PNNS into H.265.
• Demonstrated average gains 0.99% higher than prior neural network-based methods.
• Showcased the ability of PNNS to model a wide range of complex textures effectively.
• Confirmed that fully-connected networks excel in small blocks, while convolutional networks perform better in large, textured blocks.

Conclusions:

• PNNS offers a superior approach to intra image prediction compared to traditional H.265 modes.
• The adaptive nature of PNNS allows for efficient prediction across various block sizes and complexities.
• PNNS provides significant performance improvements for video compression standards.