Deep Learning Image Segmentation Based on Adaptive Total Variation Preprocessing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image segmentation is crucial for analyzing complex visual data.
• Traditional methods struggle with intricate structures and noisy backgrounds.
• Deep learning offers powerful segmentation but can be sensitive to initial image quality.

Purpose of the Study:

• To develop a robust two-stage image segmentation method.
• To improve segmentation accuracy for images with complex backgrounds and structures.
• To combine the strengths of regularization-based smoothing and deep learning.

Main Methods:

• Introduced an anisotropic regularization term with an adaptive weighted matrix in the first stage.
• Utilized the alternating direction method of multipliers (ADMMs) for convex optimization.
• Applied deep learning for segmentation on the smoothed image from the first stage.

Main Results:

• The adaptive weighting matrix effectively diffuses curves along object feature tangents.
• Complex background interference was significantly reduced.
• The two-stage method outperformed traditional and existing deep learning approaches in evaluation metrics.

Conclusions:

• The proposed two-stage method achieves high perceptual quality in image segmentation.
• It demonstrates superior performance for segmenting images with complex structures and backgrounds.
• This approach offers a promising direction for advanced image analysis tasks.