Unified DeepLabV3+ for Semi-Dark Image Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Semantic segmentation is crucial for accurate visual perception but faces challenges in dynamic scene classification.
• Existing ResNet-based DeepLabV3+ models struggle with biased filter masks, limited representational power, and loss of spatial relationships.

Purpose of the Study:

• To address the limitations of DeepLabV3+ for semantic segmentation.
• To introduce an improved model, Unified DeepLabV3+, and a novel evaluation metric, S3core.

Main Methods:

• Introduced additional dilated convolution layers with customized dilation rates to mitigate biased filter mask exploitation.
• Incorporated non-linear group normalization shortcuts to enhance representational power, particularly for semi-dark images.
• Utilized geometrically bunched pixel cues to preserve global and local spatial context.

Main Results:

• Unified DeepLabV3+ demonstrated a 3% improvement in class-wise pixel accuracy over DeepLabV3+ on the CamVid dataset.
• The proposed S3core metric, a weighted combination of pixel accuracy, IoU, and Mean BFScore, showed superior performance.
• The enhanced model proved effective for autonomous vehicles and robotics in outdoor environments.

Conclusions:

• Unified DeepLabV3+ offers a proficient solution for accurate visual perception tasks.
• The S3core metric provides a robust evaluation criterion for semantic segmentation.
• The proposed methods show significant applicability for real-world autonomous systems.