One-Shot Simple Pattern Detection without Pre-Training and Gradient-Based Strategy.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• One-shot object detection is crucial for new projects with minimal training data.
• Existing methods often require extensive pre-training and large datasets.
• A need exists for adaptable models that can learn from single images.

Purpose of the Study:

• To develop a one-shot object detection model requiring no pre-training.
• To enable automatic architecture determination for new target detection tasks.
• To detect the presence and location of simple targets with limited data.

Main Methods:

• Proposed a one-shot simple target detection model.
• Implemented a non-training parameter-gaining strategy.
• Utilized correlation coefficient-based feedforward and activation functions.
• Tested on custom project data, Brown-Yosemite, and MNIST datasets.

Main Results:

• Achieved up to 87.04% Intersection over Union (IOU) on custom project data.
• Reached 80.28% accuracy on the Brown-Yosemite dataset.
• Obtained up to 89.4% accuracy on a subset of the MNIST dataset.
• Demonstrated successful target area return and location output.

Conclusions:

• The proposed model effectively addresses one-shot object detection challenges with minimal data.
• Its automatic architecture decision and no pre-training requirements offer flexibility for new projects.
• The model's performance on diverse datasets validates its practical applicability.