Optimization of Graffiti Detection in Autonomous Drones: Effect of Image Resolution and Confidence Threshold Tuning

Not only does graffiti cleanup costs above $12 billion annually in the United States, it uses over 25,000 tons of carbon globally, requiring large trucks, cranes, and human labor to manually remove illegal graffiti. This shows the need for better and more efficient technology to reduce the cost. Although recently there is an introduction of using drones to clear graffiti, it depends on human control, requires a long time, and requires constant refilling of paint. This study designs an optimized graffiti detection machine learning model to be potentially used in graffiti removing drones and traffic cameras. We utilized transfer learning techniques by importing a pretrained YOLOv8 model to train on pre-processed dataset. To optimize the execution time and the average Intersection over Union (IoU) metrics of YOLOv8, we studied how image size and confidence threshold each affect the performance of the model. We found that the optimal image sidelength is 384 pixels and the average IoU increases with decreasing confidence threshold.