Simulation-Based Object Detector for Real-Time Defect Recognition on Industrial Infrastructures

The effective use of synthetic data for the training of recognition algorithms is considered as a major turning point in computer vision. The cost associated with the gathering of training data currently limits the use of deep learning solutions for object recognition to a selected few applications. Despite the advances in photorealistic rendering capabilities, the lack of transferability between synthetic and real training domains still constitutes a major setback. Leveraging this issue is of essence as it would enable the use of artificial images in the framework of any application able to be modeled. In this work, an automated synthetic images generation pipeline is used to train a detection model based on the YOLOv5 architecture for the real-time recognition of defects in pipe infrastructures. The effectiveness of detection is assessed on real recordings of damaged tubes and for different training configurations. The suitability of the synthetic data generation pipeline is evaluated and recommendations for future applications are drawn.