Fully Connected Network-Based Intra Prediction for Image Coding.

Area of Science:

• Computer Vision
• Machine Learning
• Video Compression

Background:

• Traditional intra prediction methods rely on fixed rules, limiting prediction accuracy.
• Existing methods often use limited contextual information from neighboring pixels.

Purpose of the Study:

• To develop a deep learning-based intra prediction method for improved video compression efficiency.
• To leverage contextual information from multiple reference lines for enhanced prediction.

Main Methods:

• A fully connected deep learning network is proposed to learn an end-to-end mapping for intra prediction.
• The network utilizes multiple reference lines, incorporating more contextual information than traditional single-line methods.

Main Results:

• The proposed deep learning method achieves an average of 3.4% bitrate saving compared to High Efficiency Video Coding (HEVC) reference software HM-16.9.
• Significant bitrate savings of 4.5% on average for 4K sequences were observed, with a maximum saving of 7.4%.

Conclusions:

• The deep learning method effectively enhances intra prediction accuracy and bitrate efficiency in video coding.
• The proposed network demonstrates good generalization ability across different bitrate settings, performing well even when trained on specific bitrates.