Explored Normalized Cut With Random Walk Refining Term for Image Segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• The Normalized Cut (NCut) model is widely used for image segmentation but struggles with excessive normalization, hindering small object and twig segmentation.
• Existing NCut models often fail to adequately segment fine structures or small salient regions due to normalization issues.

Purpose of the Study:

• To develop an improved image segmentation model that addresses the limitations of the Normalized Cut (NCut) model.
• To enhance the segmentation of small objects and fine structures like twigs.

Main Methods:

• Proposing the Explored Normalized Cut (ENCut) model, incorporating a meaningful-loop and k-step random walk for graph balancing.
• Introducing a Random Walk Refining Term (RWRT) with unsupervised random walks to add local attention and improve twig segmentation.
• Developing a move-making based strategy for efficient model solving.

Main Results:

• The ENCut model effectively reduces the energy of small salient regions, leading to enhanced small object segmentation.
• The addition of RWRT significantly improves the segmentation accuracy of thin structures like twigs.
• Experimental results demonstrate state-of-the-art performance compared to other NCut-based segmentation models on standard datasets.

Conclusions:

• The proposed ENCut model with RWRT offers a significant advancement in graph-based image segmentation.
• The model successfully overcomes the excessive normalization problem of NCut, improving segmentation for challenging objects.
• ENCut provides a robust and efficient solution for accurate image segmentation, particularly for small objects and fine details.