Anomaly detection on bridges with Computer Vision

This project was developed during a Computer Vision course at UFSC, in 2023.

The work was originally published on Github on this link.

Visual inspection plays a crucial role in assessing the condition of bridges and viaducts, being essential for maintenance, recovery, and reinforcement plans. The inevitable degradation of structures makes the adoption of evaluations relevant, transcending diagnosis and mitigating subjectivity through computer vision.

The dacl10k database, composed of 9,920 images of real inspections, categorizes 12 damage classes and 6 bridge components. This study focuses on common damages, resulting in 2,676 images of exposed reinforcements and cracks, highlighting 2,000 instances of exposed reinforcements and 3,626 of cracks.

The methodology employs convolutional neural networks for segmentation, using single-stage (YOLOv8) and two-stage (Detectron2) models with transfer learning from the COCO dataset. The data subset is divided into 80% for training and 20% for validation. Augmentation techniques such as brightness, contrast, saturation, and mirroring are applied, along with specific hyperparameters.

Results highlight the performance of YOLOv8, with 35.42% and 12.63% for mAP@50 and mAP@50:95, while Detectron2 achieves 36.64% and 13.61%. Regarding frames per second (FPS), YOLOv8 reaches 20FPS, outperforming Detectron2 with an average of 15FPS.

Both models present similar metrics when standardized with the same hyperparameters and optimization techniques. However, in the qualitative analysis of inferences, YOLOv8 proved to be more robust. The sensitivity of Detectron2's hyperparameters was evaluated, without observing significant improvements. For future work, manual filtering of the image subset is planned to select more reliable data within the application domain perspective.