Augmentation of Laser Welding Dataset through a combination of Evolutionary Optimization and Deep Learning
Gleb Solovev, Mikhail Sokolov, Aveen Hussein, and 1 more author
In Proceedings of the Genetic and Evolutionary Computation Conference Companion, NH Malaga Hotel, Malaga, Spain, 2025
In this paper, we propose a method for synthetically generating industrial data based on evolutionary approaches. The method involves applying evolutionary operators to the coordinates of a polygon describing an object’s shape and then transferring the original image style to the resulting polygon geometry.The proposed approach was validated for the laser welding problem. An augmented dataset expanded from 120 original weld images to 4500 synthetic examples, enabling the training of a convolutional neural network for weld segmentation.segmentation.