Honey Bee Conservation using Deep Learning

Develop a novel deep learning-based solution for honey bee colony monitoring, achieving 93% accuracy in detecting and classifying honey cells and brood from smartphone images, with potential applications in beekeeping and research.

Key takeaways

The researchers developed a neural network-based solution for honey bee colony monitoring using deep learning.
The task is to detect and classify honey cells and brood (bee young) in images taken by smartphone cameras.
The team used a pipeline composed of image acquisition, pre-processing, training, and testing to achieve an accuracy of 93%.
Five different architectures were tried: Convolutional Neural Networks (CNN), Fully Convolutional Networks (FCN), UNet, DenseNet, and Inception-V3.
The MobileNet was found to be the most suitable due to its efficiency and accuracy.
Challenges included dealing with lens distortion, illumination variations, and the difficulty in manual annotation of images.
A post-processing step using a binary threshold was applied to segment out the cells from the background.
F1-score and accuracy loss were used to evaluate the performance of the models.
For real-world applications, a simple software solution was developed to automatically generate reports on the health of honeybee colonies.
The team plans to open-source the software and data set, potentially enabling beekeepers and research institutions to use the technology for colony management and research.

Honey Bee Conservation using Deep Learning

More talks