Lightweight semantic segmentation network with configurable context and small object attention.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Current semantic segmentation algorithms struggle with encoding feature distortion and losing small object details.
• Existing context information exchange methods have fixed spatial ranges, limiting their effectiveness.
• High-resolution feature maintenance aids small object detection but reduces network speed.

Purpose of the Study:

• To propose a novel lightweight semantic segmentation network, CCSONet, designed to overcome the limitations of existing methods.
• To enhance feature representation by incorporating configurable context and attention mechanisms for small objects.
• To improve both accuracy and efficiency in semantic segmentation tasks.

Main Methods:

• Developed a lightweight semantic segmentation network named CCSONet.
• Introduced a long-short distance configurable context feature enhancement module (LSCFEM) for flexible spatial context.
• Implemented a small object attention decoding module (SOADM) to focus on and enhance small object features.

Main Results:

• CCSONet achieved 76.9 mIoU on the Cityscapes dataset and 73.1 mIoU on the Camvid dataset.
• The network maintained high operational speeds of 87 FPS and 138 FPS on Cityscapes and Camvid, respectively.
• CCSONet demonstrated superior accuracy compared to other lightweight semantic segmentation algorithms.

Conclusions:

• CCSONet effectively addresses feature distortion and small object loss in semantic segmentation.
• The proposed LSCFEM and SOADM modules provide configurable context and targeted attention for improved performance.
• CCSONet offers a promising solution for efficient and accurate lightweight semantic segmentation.