A Fast Attention-Guided Hierarchical Decoding Network for Real-Time Semantic Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Semantic segmentation is crucial for scene understanding but often suffers from slow inference speeds due to complex models.
• Real-time applications require efficient semantic segmentation models that balance accuracy and speed.

Purpose of the Study:

• To propose a novel network, FAHDNet, for fast and accurate real-time semantic segmentation.
• To address the trade-off between accuracy and inference speed in existing semantic segmentation models.

Main Methods:

• Developed an asymmetric U-shaped network (FAHDNet) featuring an encoder with a multi-scale bottleneck residual unit (MBRU) incorporating attention and decomposition convolution.
• Introduced a spatial information compensation (SIC) module to recover lost spatial texture information.
• Implemented a global attention (GA) module in the decoder to enhance feature interaction and a lightweight hierarchical decoder for multi-scale feature integration.

Main Results:

• FAHDNet achieved 70.6% mIoU at 135 FPS on the Cityscapes dataset.
• The network reached 67.2% mIoU at 335 FPS on the Camvid dataset.
• Demonstrated superior performance compared to existing networks in terms of both accuracy and inference speed.

Conclusions:

• FAHDNet effectively balances high accuracy with fast inference speeds for real-time semantic segmentation.
• The proposed network enhances practical usability in applications demanding rapid scene understanding.
• The integration of attention mechanisms and hierarchical decoding contributes to improved multi-scale object segmentation.