Multipattern Mining Using Pattern-Level Contrastive Learning and Multipattern Activation Map.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Visual patterns are fundamental to image understanding and computer vision tasks.
• Existing methods often assume a single pattern per category, failing to capture the complexity of many-to-one mappings.
• Discovering and distinguishing varied patterns within a single category remains a significant challenge.

Purpose of the Study:

• To propose a novel framework, PaclMap, for mining multiple visual patterns within image categories.
• To address the limitations of existing methods in handling many-to-one pattern mappings.
• To develop a method that can discover both discriminative and frequent patterns.

Main Methods:

• PaclMap framework utilizes medium-grained features to represent visual patterns.
• Employs unsupervised pattern-level contrastive learning.
• Incorporates a multi-pattern activation map for joint optimization.

Main Results:

• PaclMap demonstrated superior performance across four benchmark datasets (Place-20, ILSVRC-20, VOC, and Travel).
• Achieved an average improvement of 2.9% in accuracy compared to state-of-the-art methods.
• Showcased an average improvement of 12.3% in pattern frequency discovery.

Conclusions:

• PaclMap effectively mines and distinguishes multiple visual patterns within image categories.
• The proposed framework offers a significant advancement in pattern mining for computer vision.
• This approach enhances the ability to understand complex visual regularities in images.