Efficient In-loop Filtering Based on Enhanced Deep Convolutional Neural Networks for HEVC.

Area of Science:

• Computer Vision
• Digital Signal Processing
• Machine Learning

Background:

• High Efficiency Video Coding (HEVC) compression introduces artifacts that degrade video quality.
• Current in-loop filtering methods in HEVC struggle to fully eliminate these artifacts.
• Deep learning shows promise for enhancing video processing tasks.

Purpose of the Study:

• To propose an efficient in-loop filtering algorithm for HEVC using enhanced deep convolutional neural networks (EDCNN).
• To address limitations in traditional convolutional neural network models for video artifact removal.
• To significantly improve the performance of in-loop filtering in HEVC compression.

Main Methods:

• Analysis of traditional convolutional neural network (CNN) issues: normalization, learning ability, and loss functions.
• Development of EDCNN incorporating a weighted normalization method, feature information fusion block, and a precise loss function.
• Evaluation using Peak Signal-to-Noise Ratio (PSNR) enhancement and smoothness, Rate-Distortion (RD) performance, subjective testing, and computational complexity.

Main Results:

• The proposed EDCNN algorithm demonstrates superior artifact elimination compared to the standard HM16.9 filter.
• Achieved an average of 6.45% BDBR reduction, indicating improved compression efficiency.
• Obtained 0.238 dB BDPSNR gains, signifying enhanced video quality.

Conclusions:

• The EDCNN offers an effective solution for artifact reduction in HEVC compressed videos.
• The proposed enhancements to CNN models lead to better video quality and compression performance.
• EDCNN presents a promising approach for advanced in-loop filtering in video coding standards.