Rethinking 1D convolution for lightweight semantic segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Existing lightweight semantic segmentation networks (LSNet) suffer from low precision and high parameter counts.
• There is a growing need for efficient semantic segmentation models deployable on resource-constrained devices.

Purpose of the Study:

• To design a novel lightweight semantic segmentation network (LSNet) that overcomes the limitations of existing models.
• To improve segmentation accuracy and reduce parameter count for efficient deployment on tiny devices.

Main Methods:

• Developed a full 1D convolutional LSNet incorporating three key modules: 1D multi-layer space (1D-MS), 1D multi-layer channel (1D-MC), and flow alignment (FA).
• Integrated global feature extraction using 1D convolutions within 1D-MS and 1D-MC modules.
• Designed a 1D-mixer encoder based on transformer architecture for fusion encoding and an attention pyramid with FA (AP-FA) for feature decoding and alignment.

Main Results:

• Achieved 72.6 mIoU and 95.6 FPS on Cityscapes, and 70.5 mIoU and 122 FPS on CamVid.
• Demonstrated successful deployment on mobile devices with a latency of 224 ms (trained on ADE2K dataset).
• The proposed LSNet has only 0.62 M parameters, offering the highest segmentation accuracy within 1 M parameters compared to state-of-the-art methods.

Conclusions:

• The designed LSNet achieves a superior balance between segmentation accuracy and parameter efficiency.
• The network's performance and low latency confirm its applicability for real-world applications on mobile and tiny devices.
• The novel architecture effectively fuses multi-level features and enhances global information extraction.