Fast detection of bees using deep learning and bayesian optimization
In commercial beekeeping, monitoring the apiaries is difficult as they are often spread over large distances. Building a vision-based hive monitoring system is a promising—albeit difficult—solution to this problem. In this thesis, I approach this task by implementing and training three deep learning based object detection models to detect bees and predators: fast region-based convolutional neural networks (Fast R-CNN), You Only Look Once version 2 (YOLO2), and the single shot multi-box detector (SSD). I also use Bayesian optimization (BO) to tune the detector hyperparameters and test whether it is effective for this task. After training the models and tuning their hyperparameters by hand, I obtained a best F1 score of 0.443 for Fast R-CNN, 0.306 for YOLO2, and 0.428 for SSD on a test dataset. After tuning the hyperparameters using BO, I obtained scores of 0.442 for Fast R-CNN, 0.317 for YOLO2, and 0.459 for SSD.
Computer vision, Object detection, Machine learning, Bees