Degraded Image Semantic Segmentation with Dense-Gram Networks.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Degraded image semantic segmentation is vital for safety-critical applications like autonomous driving.
• Image degradations significantly reduce semantic segmentation accuracy, creating a gap between clean and degraded image feature distributions.
• Existing methods to bridge this gap include pre-processing, joint training, and fine-tuning, but they have limitations.

Purpose of the Study:

• To propose a novel network, the Dense-Gram Network, for effectively segmenting degraded images.
• To address the performance gap caused by feature distribution differences between clean and degraded images.
• To achieve state-of-the-art performance in degraded image semantic segmentation.

Main Methods:

• Developed a novel Dense-Gram Network architecture.
• Trained and evaluated the network on degraded images synthesized from PASCAL VOC 2012, SUNRGBD, CamVid, and CityScapes datasets.
• Compared the Dense-Gram Network against conventional strategies for handling degraded image data.

Main Results:

• The proposed Dense-Gram Network significantly reduces the gap between clean and degraded image feature distributions.
• Achieved state-of-the-art semantic segmentation performance on various degraded image datasets.
• Demonstrated superior effectiveness compared to conventional methods.

Conclusions:

• The Dense-Gram Network offers a more effective solution for degraded image semantic segmentation.
• This advancement is critical for improving the reliability of autonomous driving and navigation systems.
• The proposed method sets a new benchmark for segmenting degraded visual data.