Bio-inspired contour extraction via EM-driven deformable and rotatable directivity-probing mask

Area of Science:

• Computer Vision
• Image Processing
• Computational Neuroscience

Background:

• Traditional contour extraction methods often rely on segmentation or explicit linking mechanisms.
• Human perception groups edges based on Gestalt principles, which are challenging to replicate computationally.
• Existing algorithms struggle with noise, texture, and low-contrast boundaries.

Purpose of the Study:

• To develop a bio-inspired, edge-oriented approach for unsupervised perceptual contour extraction.
• To identify edge points that naturally group into object boundaries, mimicking human perception.
• To improve contour extraction resilience against noise and texture, and enhance detection of low-contrast edges.

Main Methods:

• A novel bio-inspired, edge-oriented approach using a dynamic mask and Expectation-Maximization (EM) algorithm.
• The dynamic mask statistically assesses inter-edge relations and identifies principal directions for edge grouping.
• Utilizes EM-based Bayesian inference for iterative mask deformation and rotation to fit pixel likelihoods.

Main Results:

• The proposed method effectively identifies contiguous edges belonging to the same object while suppressing noise and textures.
• Achieved contours align well with Gestalt laws of continuity, similarity, and proximity.
• Demonstrated superiority over existing methods in tracking curved contours, noise/texture resilience, and detecting low-contrast contours.

Conclusions:

• The novel bio-inspired approach offers a robust and effective solution for perceptual contour extraction.
• The dynamic mask and EM-based Bayesian inference provide a powerful framework for unsupervised edge grouping.
• This method significantly advances the state-of-the-art in contour detection, particularly in challenging image conditions.