Faster SCDNet: Real-Time Semantic Segmentation Network with Split Connection and Flexible Dilated Convolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Semantic segmentation is crucial for realistic scenarios.
• Dense connections in backbone networks improve accuracy but reduce inference speed.

Purpose of the Study:

• To propose SCDNet, a backbone network balancing speed and accuracy in semantic segmentation.
• To enhance gradient propagation efficiency and network inference speed.

Main Methods:

• Introduced a streamlined, lightweight backbone with a split connection structure for increased inference speed.
• Utilized flexible dilated convolutions with varying dilation rates for richer receptive fields.
• Implemented a three-level hierarchical module for balancing multi-resolution feature maps.
• Employed a refined, flexible, and lightweight decoder.

Main Results:

• Achieved a trade-off between accuracy and speed on Cityscapes and Camvid datasets.
• Demonstrated a 36% improvement in frames per second (FPS).
• Obtained a 0.7% improvement in mean intersection over union (mIoU) on the Cityscapes test set.

Conclusions:

• SCDNet offers a viable solution for semantic segmentation tasks requiring both high accuracy and fast inference.
• The proposed architectural components effectively address the speed-accuracy limitations of existing models.