Genetic Programming With a New Representation to Automatically Learn Features and Evolve Ensembles for Image...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image classification is challenging due to high intra-class variation.
• Ensemble methods improve classification accuracy but require accurate and diverse base classifiers.
• Effective feature extraction, crucial for image classification, often necessitates domain expertise.

Purpose of the Study:

• To propose an evolutionary approach using genetic programming for automatic feature extraction and ensemble learning in image classification.
• To develop a method that simultaneously learns informative features and evolves effective ensembles from raw image data.
• To address the need for domain knowledge in feature extraction and the diversity requirement in ensemble methods.

Main Methods:

• A novel evolutionary approach based on genetic programming is introduced.
• The approach utilizes a new individual representation, function set, and terminal set for effective problem-solving.
• It takes raw images as input and outputs class label predictions via evolved classifiers.

Main Results:

• The proposed method automatically extracts informative features from raw images.
• It effectively addresses the diversity challenge in ensemble learning.
• The approach automatically selects and optimizes parameters for ensemble classification algorithms.
• Evaluated on 13 diverse image classification datasets, the method demonstrated superior classification accuracy compared to numerous existing approaches.

Conclusions:

• The evolutionary approach successfully evolves accurate and diverse solutions for image classification.
• It automates both feature extraction and ensemble optimization, reducing the need for domain expertise.
• The method offers a powerful and versatile tool for tackling complex image classification tasks.