Binary Image Classification: A Genetic Programming Approach to the Problem of Limited Training Instances.

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning

Background:

• Image classification is a challenging task in computer vision and pattern recognition.
• Developing models that mimic human visual learning with few instances is difficult.
• Existing methods struggle with learning new classes from limited data.

Purpose of the Study:

• To investigate the performance, robustness, and complexity of two novel genetic programming (GP) methods: one-shot GP and compound-GP.
• To evaluate these GP methods for binary image classification using limited instances per class.
• To compare the proposed methods against other GP and non-GP approaches.

Main Methods:

• Evolving programs for binary image classification using one-shot GP and compound-GP.
• Utilizing ten datasets of varying difficulty for evaluation.
• Comparing performance against two other GP methods and six non-GP methods.

Main Results:

• One-shot GP and compound-GP achieved superior or comparable performance to existing methods.
• The features extracted by these GP methods improved the performance of other classifiers.
• The evolved programs showed competitive robustness and complexity.

Conclusions:

• One-shot GP and compound-GP are effective for image classification, particularly with limited data.
• These methods offer a promising approach for developing efficient and robust image classification models.
• The extracted features have broad applicability in enhancing various classification tasks.