Shearlets as feature extractor for semantic edge detection: the model-based and data-driven realm.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Semantic edge detection is crucial for image analysis, extracting vital information from image edges.
• Existing methods face challenges due to the distinct abstraction levels of edge detection and classification, known as the distracted supervision paradox.
• This paradox limits the performance of purely supervised models in semantic edge detection.

Purpose of the Study:

• To develop a novel hybrid method for semantic edge detection that overcomes the distracted supervision paradox.
• To achieve high performance in semantic edge detection tasks.
• To demonstrate the method's efficiency in terms of parameter count compared to data-driven approaches.

Main Methods:

• A hybrid approach combining the model-based shearlet transform with a data-driven convolutional neural network (CNN).
• Shearlets provide optimally sparse approximations for image models.
• A specifically designed CNN is integrated with shearlet properties.

Main Results:

• The proposed hybrid method successfully avoids the distracted supervision paradox.
• Achieved high performance in semantic edge detection.
• Demonstrated significantly fewer parameters compared to pure data-driven methods.
• Outperformed previous methods in applications like tomographic reconstruction.

Conclusions:

• The novel hybrid method offers a significant advancement in semantic edge detection.
• This approach provides a more efficient and effective solution compared to existing methods.
• Highlights the potential of hybrid methods in biomedical imaging and other fields.