A Dual-Channel and Frequency-Aware Approach for Lightweight Video Instance Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video instance segmentation is crucial for intelligent sensing in applications like automated surveillance and robotics.
• Traditional methods struggle with high computational costs and slow speeds, especially in resource-limited settings.
• Efficient real-time target tracking is essential for accurate perception in dynamic environments.

Purpose of the Study:

• To propose a Dual-Channel and Frequency-Aware Approach for Lightweight Video Instance Segmentation (DCFA-LVIS).
• To enhance feature extraction and representation capabilities for improved accuracy.
• To reduce model complexity and improve segmentation efficiency for practical deployment.

Main Methods:

• A DCEResNet backbone with a dual-channel feature enhancement mechanism for improved accuracy.
• A dual-frequency perceptual enhancement network utilizing an independent instance query mechanism.
• A frequency-aware attention mechanism to capture high and low-frequency instance features.

Main Results:

• The DCFA-LVIS model achieves state-of-the-art segmentation performance on the YouTube-VIS dataset.
• The proposed method demonstrates high efficiency with a significantly reduced number of parameters.
• Experimental results validate the model's practicality and effectiveness.

Conclusions:

• DCFA-LVIS offers an efficient and accurate solution for video instance segmentation.
• The approach significantly enhances the application efficiency and adaptability of intelligent sensing technologies.
• This method provides strong support for the widespread deployment of visual perception in video data processing.