Genetic Programming-Based Discriminative Feature Learning for Low-Quality Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Evolutionary Computation

Background:

• Learning discriminative features from low-quality images is crucial for applications like autonomous driving and surveillance.
• Image distortions such as blur, low contrast, and noise, along with variations in scale, rotation, illumination, and viewpoint, pose significant challenges.
• Traditional image preprocessing often requires manual intervention and domain expertise.

Purpose of the Study:

• To propose a novel evolutionary learning approach using genetic programming (GP), termed EFLGP, for automatic feature extraction from low-quality images.
• To enable robust image classification without human intervention in preprocessing.
• To develop a method that effectively handles image variations and distortions.

Main Methods:

• Developed a new genetic programming approach (EFLGP) with a novel program structure, function set, and terminal set.
• Incorporated common image preprocessing operators as functions within the GP framework.
• EFLGP is designed to detect small image regions, apply operators, and extract a flexible number of discriminative features.

Main Results:

• EFLGP demonstrated significantly better or comparable performance across eight diverse datasets compared to benchmark methods using handcrafted and deep features.
• The proposed method achieved high accuracy on images with blur, low contrast, and noise.
• EFLGP exhibited greater invariance to image degradations than benchmark approaches.

Conclusions:

• The evolutionary learning approach using genetic programming (EFLGP) effectively learns discriminative features from low-quality images.
• EFLGP offers an automated solution for image preprocessing and feature extraction, reducing reliance on human expertise.
• The method shows promising robustness and invariance to common image distortions.