VICTOR ESCRIBANO GARCIA'S PORTFOLIO
About the project
I created a synthetic pipe crack generator using Blender, designed to produce realistic images of pipe sections with randomly generated cracks. Each iteration varies camera parameters, illumination conditions, and other factors to ensure a diverse and comprehensive dataset. This innovative tool automatically labels the cracks, making it an excellent resource for generating precise datasets needed for segmentation tasks.
The dataset generated by this tool includes 2000 synthetic images and 900 real images, providing a robust foundation for training machine learning models. The synthetic images simulate a wide range of conditions, while the real images add authenticity and validation. This combination ensures that the trained model can effectively handle various scenarios and improve its accuracy in crack detection.
Utilizing the YOLOv8 medium-size model, the generated dataset demonstrates impressive performance in detecting cracks not only in pipes but also on ground surfaces. The model's ability to generalize and excel in different crack detection tasks showcases the robustness and versatility of the system, making it a valuable asset for various industrial applications and research initiatives.