Refined UNet v3: Efficient end-to-end patch-wise network for cloud and shadow segmentation with multi-channel...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Semantic segmentation is crucial for computer vision but struggles with precise edge detection due to limitations in modeling pixel relationships.
• Existing methods like Refined UNet v2 face challenges with ambiguous edges and computational efficiency in their Conditional Random Field (CRF) layers.

Purpose of the Study:

• To introduce Refined UNet v3, an improved model addressing the limitations of Refined UNet v2 for edge-precise semantic segmentation.
• To enhance the bilateral message-passing kernel and Gaussian filtering implementation within the CRF layer for better performance and efficiency.

Main Methods:

• Refined UNet v3 utilizes a UNet backbone for coarse segmentation and an embedded CRF layer for edge refinement.
• The CRF layer incorporates a multi-channel guided Gaussian filter for improved detection of ambiguous edges.
• Fast Fourier Transform (FFT)-based Gaussian filtering is employed for efficient and potentially range-agnostic message passing.

Main Results:

• Refined UNet v3 demonstrates superior performance in detecting vague edges, shadow retrieval, and handling isolated redundant regions compared to its predecessor.
• The model shows practical efficiency in its TensorFlow implementation, accelerating the message-passing procedure.
• The enhanced model proves effective for segmentation on multi-spectral datasets, confirming advancements in shadow retrieval.

Conclusions:

• Refined UNet v3 represents a significant advancement in edge-precise semantic segmentation, offering improved accuracy and efficiency.
• The model's enhancements, particularly in the CRF layer's message-passing mechanism, effectively address challenges with ambiguous edges.
• Refined UNet v3 is a versatile and efficient tool for various computer vision tasks, including remote sensing applications.