MFAFNet: A Lightweight and Efficient Network with Multi-Level Feature Adaptive Fusion for Real-Time Semantic...

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Real-time semantic segmentation is crucial for resource-constrained applications like autonomous driving.
• Existing methods often struggle with receptive field size and detail preservation, impacting accuracy.
• There is a need for efficient networks balancing speed, accuracy, and model size.

Purpose of the Study:

• To propose a lightweight and efficient Multi-level Feature Adaptive Fusion Network (MFAFNet) for real-time semantic segmentation.
• To address the limitations of current networks in achieving large receptive fields and preserving low-level details.
• To optimize inference speed without sacrificing segmentation accuracy.

Main Methods:

• Designed a separable asymmetric reinforcement non-bottleneck module for extracting short- and long-range contextual information.
• Developed a feature adaptive fusion module to balance multi-resolution feature maps and reduce spatial detail loss.
• Utilized optimized convolutions to enhance inference speed.

Main Results:

• MFAFNet achieved 75.9% mIoU on Cityscapes and 69.9% mIoU on Camvid datasets.
• The network operates at 60.1 FPS (Cityscapes) and 82.6 FPS (Camvid) without pre-training or post-processing.
• MFAFNet has only 1.27 million parameters, demonstrating a compact model size.

Conclusions:

• The proposed MFAFNet offers an excellent trade-off between inference speed, segmentation accuracy, and model size.
• The network is suitable for real-time semantic segmentation in demanding applications.
• The novel modules effectively capture contextual information and preserve spatial details.