Deep Dual-Resolution Road Scene Segmentation Networks Based on Decoupled Dynamic Filter and Squeeze-Excitation Module

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Image semantic segmentation is crucial for autonomous driving assistance.
• Challenges include complex road scenes and real-time processing demands.

Purpose of the Study:

• To propose a novel network, Deep Dual-resolution Road Scene Segmentation Networks (DDF&SE-DDRNet), to address segmentation challenges.
• To enhance segmentation accuracy and efficiency for autonomous driving applications.

Main Methods:

• The DDF&SE-DDRNet incorporates a decoupled dynamic filter to reduce parameters and enable dynamic weight adjustment of convolution kernels.
• Integration of Squeeze-and-Excitation modules allows local feature maps to acquire global features, mitigating local image interference.

Main Results:

• Experimental results on the Cityscapes dataset demonstrate a segmentation accuracy improvement of at least 2% compared to existing algorithms.
• The DDF&SE-DDRNet achieves a satisfactory inference speed, suitable for real-time applications.

Conclusions:

• The proposed DDF&SE-DDRNet effectively enhances road scene segmentation accuracy and efficiency.
• This network presents a promising solution for the challenges in autonomous driving assistance technology.