A Multi-Module Fusion Framework for Restoring Human and Machine Vision Quality in Compressed Video

Area of Science:

• Computer Vision
• Multimedia Systems
• Image Processing

Background:

• Video compression introduces artifacts, degrading visual quality and machine task performance.
• This is a critical issue for sensor-based systems like surveillance cameras and mobile devices.
• Existing methods struggle to balance perceptual quality and task-specific performance.

Purpose of the Study:

• To propose a joint human-machine video enhancement framework for compressed videos.
• To simultaneously improve human perceptual quality and machine vision task performance.
• To address the limitations of current compressed video enhancement techniques.

Main Methods:

• A novel framework integrating four complementary modules: Spatio-Temporal Fusion, High-Frequency Semantic Fusion, Texture-Guided Model, and Refined Attention Residual Quality Enhancement.
• Leveraging inter-frame correlations and recovering structurally important details for machine tasks.
• Adaptively emphasizing salient regions to enhance low-level visual features.

Main Results:

• The proposed framework consistently outperforms existing methods in compressed video enhancement.
• Achieved higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics.
• Demonstrated improved performance in downstream machine vision tasks like object detection and video object segmentation.

Conclusions:

• The joint human-machine framework effectively restores compressed video content while preserving task-relevant semantics.
• The method shows practical applicability for compressed video enhancement in sensor-based systems.
• Highlights potential for intelligent surveillance and autonomous imaging platforms.