Summary

This thesis project presents a novel approach for detecting corrosion using drone-based surveillance systems by modifying the YOLOv5 object detection model and deploying it to a VOXL 2 flight controller for real-time image processing. The central goal is to enhance the model's accuracy in detecting corrosion instances and make the developed solution easily replicable for future use within the company.
The first part of the project involves the construction of a comprehensive dataset using drone-captured images of vapor pipes, further enriched with data augmentation techniques.
Following this, the YOLOv5 model is modified by adding a Convolutional Block Attention Module (CBAM) to better capture the fine-grained details of corrosion. The modified model is then optimized through a rigorous training and evaluation process, resulting in significantly improved performance metrics, as evidenced by the comparative analysis with the original YOLOv5 model.
The developed model is successfully deployed to the VOXL 2 flight controller, demonstrating its real-time corrosion detection capabilities in drone-captured images. Lastly, a toolkit is developed to facilitate the company's ability to train and adapt the model for future needs, ensuring the project's sustainability.
Despite the project's success, certain limitations are identified, such as dataset diversity and real-world performance evaluation, providing a direction for future research. This work contributes to the field of corrosion detection methodologies, presenting a robust, efficient, and adaptable model.