A Large Kernel Convolutional Neural Network with a Noise Transfer Mechanism for Real-Time Semantic Segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Large kernels and Atrous convolution increase receptive fields in semantic segmentation.
• Fixed kernel sizes limit adaptive multi-scale feature capture and global context utilization.

Purpose of the Study:

• To enhance semantic segmentation models by overcoming limitations of fixed kernel sizes.
• To improve adaptive multi-scale feature extraction and global contextual information leveraging.

Main Methods:

• Combining Atrous convolution with large kernel convolution using varied dilation rates.
• Implementing a dynamic selection mechanism for adaptive spatial feature highlighting.
• Proposing a Multi-Scale Contextual Noise Transfer Matrix (NTM) for supervision signal correction.

Main Results:

• LKNTNet achieves state-of-the-art speed and accuracy on benchmark datasets.
• Achieved 80.05% mIoU on Cityscapes (80.7 FPS) and 42.7% mIoU on ADE20K (143.6 FPS).

Conclusions:

• The proposed method effectively addresses multi-scale feature limitations in large kernel networks.
• LKNTNet demonstrates superior generalization capability and performance in semantic segmentation tasks.