Discriminative Shape Feature Pooling in Deep Neural Networks

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning excels at generic feature extraction from large datasets.
• Handcrafted features offer domain-specific knowledge and intuitive visual understanding.
• Integrating handcrafted and deep features faces challenges with parameter quality.

Purpose of the Study:

• To propose a method for enriching deep network features using discriminative shape information.
• To guide neural network parameter updates with explicit domain knowledge.
• To generate image representations benefiting from both handcrafted and deep learned features.

Main Methods:

• Injecting discriminative shape features (edge tokens, curve partitioning points) into deep networks.
• Adjusting the internal parameter update process of neural networks.
• Training modified neural networks with domain knowledge guidance.

Main Results:

• The proposed method effectively enriches deep network features.
• Experimental results confirm efficacy on both large and small training datasets.
• Improved performance and reduced computational costs compared to existing models.

Conclusions:

• The method successfully integrates handcrafted and deep learned features.
• The approach offers a robust solution for improving deep learning models in computer vision tasks.
• This technique enhances model performance and efficiency across various dataset sizes.