A New Regularization for Deep Learning-Based Segmentation of Images with Fine Structures and Low Contrast

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep learning excels in image segmentation but often neglects spatial pixel dependencies.
• Existing spatial regularization methods like total variation may not suit complex images with fine structures or low contrast.

Purpose of the Study:

• To develop a new regularization technique for deep learning-based image segmentation.
• To improve the connectivity and robustness of segmentation results, particularly for challenging image types.

Main Methods:

• A novel regularization method was derived to enhance segmentation connectivity.
• The proposed regularization was integrated into deep learning models.
• Its effectiveness was evaluated on both deep learning and unsupervised segmentation methods.

Main Results:

• The new regularization significantly improved segmentation performance by enhancing connectivity.
• The method effectively addressed challenges posed by low contrast and fine structures in images.
• Both deep learning and unsupervised segmentation approaches benefited from the proposed regularization.

Conclusions:

• The developed regularization technique offers a valuable improvement for image segmentation tasks.
• It enhances segmentation quality by addressing connectivity and low-contrast issues.
• The method shows broad applicability across different segmentation approaches.