Iterative Training of Neural Networks for Intra Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Image and Video Compression

Background:

• Block-based image and video codecs rely on intra prediction for efficient compression.
• Existing intra prediction methods have limitations in adapting to diverse image content.
• Neural networks offer potential for advanced prediction but require effective training strategies.

Purpose of the Study:

• To develop an iterative training methodology for neural networks applied to intra prediction.
• To enhance the rate-distortion performance of image and video codecs.
• To investigate the impact of iterative training on prediction function learning and data cleansing.

Main Methods:

• Iterative training of neural networks using codec-partitioned image blocks and their contexts.
• Progressive retraining of neural networks with data from previous iterations.
• Integration of iteratively trained neural networks into H.265 (HM-16.15) and H.266 (VTM-5.0) codecs.

Main Results:

• The iterative training enables neural networks to learn novel intra prediction functions.
• Significant rate-distortion improvements were achieved compared to existing methods.
• A mean BD-rate reduction of -4.2% was obtained in H.265, surpassing the state-of-the-art by -1.8%.
• A mean BD-rate reduction of -1.9% was achieved in H.266.

Conclusions:

• Iterative neural network training is an effective approach for improving intra prediction in video codecs.
• The proposed method offers a significant advancement in compression efficiency.
• The iterative process facilitates essential data cleansing for robust neural network training.