Automatically Evolving Texture Image Descriptors Using the Multitree Representation in Genetic Programming Using Few...

Area of Science:

• Computer Vision
• Machine Learning
• Evolutionary Computation

Background:

• Image classification model performance relies heavily on feature extraction.
• Manual feature design is costly and requires domain expertise.
• Automating feature design can reduce costs and effort.

Purpose of the Study:

• To automatically evolve an image descriptor using genetic programming (GP).
• To address the challenge of feature design in image classification, particularly in few-shot learning scenarios.
• To reduce the reliance on costly domain experts for feature engineering.

Main Methods:

• Utilized genetic programming (GP) with a multitree program representation to automatically evolve image descriptors.
• Operated directly on raw pixel values to generate feature vectors.
• Adapted seven datasets to a few-shot setting (two instances per class) for evaluation.

Main Results:

• The evolved image descriptor significantly outperformed six handcrafted descriptors, one evolutionary computation-based descriptor, and three convolutional neural network (CNN) based methods.
• The method demonstrated strong performance in few-shot image classification settings.
• Analysis of evolved programs revealed identifiable patterns.

Conclusions:

• Automated feature descriptor evolution using GP is a viable and effective approach for image classification.
• The proposed method offers a cost-effective alternative to manual feature engineering.
• This technique shows promise for improving performance in data-scarce scenarios.