Deep Active Shape Model for Robust Object Fitting.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Object recognition and localization remain challenging, particularly for objects with diverse shapes and appearances.
• Deep learning (DL) has advanced the field, but integrating prior knowledge about shape and appearance is complex.
• Active Shape Models (ASM) offer a principled way to incorporate prior knowledge but can be sensitive to data quality and outliers.

Purpose of the Study:

• To propose a new Active Shape Model (ASM) framework integrating handcrafted and deep learning (DL) features.
• To enhance model fitting performance using a probabilistic framework robust to outliers and multiple observations.
• To evaluate the framework's effectiveness in facial landmark fitting and cardiac image segmentation.

Main Methods:

• Comparison of various image features, including handcrafted (HOG, SIFT, texture) and DL-based features, for observation extraction.
• Implementation of a probabilistic framework utilizing the Generalized Expectation-Maximization algorithm for outlier handling and multiple observations.
• Evaluation on facial landmark fitting and left ventricle endocardium segmentation in cardiac MRI volumes.

Main Results:

• The proposed ASM framework demonstrates robustness to outliers, adverse initialization, and large search regions.
• Combining ASM with DL-based features yields competitive performance against state-of-the-art DL approaches (FCN, U-Net, CNN Cascade).
• The framework successfully integrates the strengths of statistical models and deep learning.

Conclusions:

• The novel deep ASM probabilistic framework effectively combines statistical prior knowledge with the power of deep learning.
• This integrated approach offers a robust and competitive solution for challenging object recognition and localization tasks.
• The findings suggest a promising direction for advancing both statistical modeling and deep learning in computer vision and medical imaging.