Block Partitioning Information-Based CNN Post-Filtering for EVC Baseline Profile.

Area of Science:

• Computer Vision
• Video Compression
• Machine Learning

Background:

• The proliferation of video applications and Internet of Things (IoT) devices necessitates efficient video coding.
• The MPEG-5 Essential Video Coding (EVC) standard, particularly its Baseline profile, offers low complexity suitable for IoT but suffers from coding artifacts like blocking.
• Blocking artifacts at block boundaries are a critical issue impacting visual quality in EVC Baseline.

Purpose of the Study:

• To propose a novel post-filter to mitigate blocking artifacts in the EVC Baseline profile.
• To leverage block partitioning information and Convolutional Neural Networks (CNNs) for artifact reduction.
• To enhance the visual quality and coding efficiency of EVC for IoT video data.

Main Methods:

• Development of a post-filter utilizing a CNN.
• The CNN is trained using block partitioning information from EVC encoded video.
• Experimental evaluation comparing the proposed filter against pre-filtered video.

Main Results:

• Objective improvements in Peak Signal-to-Noise Ratio (PSNR) of approximately 0.57 dB (All-Intra) and 0.37 dB (Low-Delay).
• Significant bitrate reduction: 11.62% (AI) and 10.91% (LD) for Luma and Chroma components.
• Subjective visual quality enhancement, notably reducing blocking artifacts.

Conclusions:

• The proposed CNN-based post-filter effectively addresses blocking artifacts in EVC.
• The method offers substantial improvements in both objective metrics (PSNR) and subjective visual quality.
• This approach presents a viable solution for enhancing video coding efficiency for resource-constrained IoT environments.