Abstract Machine vision and AI-based defect detection systems are increasingly deployed in manufacturing to support consistent product quality and high production efficiency. However, these automated inspection systems often suffer from sensitivity to imaging variability, dependence on large labeled datasets, and the need for manually engineered preprocessing pipelines–limitations that hinder accuracy and reliability in real industrial conditions. This study presents a novel approach for optimizing image preprocessing strategies for neural network–based defect detection using a genetic algorithm (GA)-driven evolutionary framework. The method systematically explores a set of 48 preprocessing operations and automatically evolves optimal filter sequences through multi-objective fitness evaluations incorporating classification accuracy, computational efficiency, and preprocessing robustness. The genetic algorithm generates diverse preprocessing sequences of varying lengths (3–6 filters) and evaluates a broad range of population sizes (20–100 individuals) and generation limits (20–150 generations) to identify configurations that maximize detection performance while reducing data requirements. Extensive experiments across three product categories show that GA-optimized preprocessing significantly outperforms raw-image baselines and manually designed preprocessing pipelines. Results demonstrate substantial gains in classification accuracy (up to 15%) and improved data efficiency, requiring 30–60% fewer training images to achieve target performance. The findings confirm that evolutionary optimization provides a robust and scalable solution for industrial defect detection, enabling more reliable and efficient machine vision systems for modern manufacturing environments.
Challa et al. (Mon,) studied this question.