Deep Multiphase Level Set for Scene Parsing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Fully Convolutional Networks (FCNs) are standard for image segmentation but struggle with accurate object boundary prediction.
• Existing FCN methods often yield parsing results with imprecise boundaries.
• Level set active contours offer sub-pixel accuracy but are sensitive to initial parameters.

Purpose of the Study:

• To propose a novel Deep Multiphase Level Set (DMLS) method for semantic scene parsing.
• To address the limitations of FCNs in boundary prediction and level sets in initialization sensitivity.
• To achieve more accurate semantic scene parsing with improved boundary delineation.

Main Methods:

• The proposed Deep Multiphase Level Set (DMLS) method integrates multiphase level sets into deep neural networks.
• It comprises three modules: recurrent FCNs for multi-level image representations, an adaptive multiphase level set for discriminative contour driving, and deeply supervised learning for training.
• Recurrent FCNs capture diverse contextual information, while the adaptive level set leverages both global and local image data.

Main Results:

• The DMLS method demonstrated superior performance in semantic scene parsing tasks.
• Achieved new state-of-the-art results on three public benchmark datasets.
• The integration of recurrent FCNs and adaptive multiphase level sets significantly improved boundary accuracy.

Conclusions:

• The proposed Deep Multiphase Level Set (DMLS) method effectively enhances semantic scene parsing accuracy, particularly around object boundaries.
• DMLS overcomes limitations of traditional FCNs and level set methods by combining their strengths.
• The method represents a significant advancement in deep learning for image segmentation and scene understanding.