Bee Health Computer Vision: Varroa Mite Detection
This project developed an AI-based solution for monitoring bee health, with a focus on
distinguishing between bees and wasps and detecting varroa mites on bees. The initiative
addressed the environmental challenge of maintaining healthy bee populations, given the
detrimental impact of varroa mites.
A ResNet RS 350 model, fine-tuned on TensorFlow's bee_dataset, formed the core of this
system. The model's capabilities were twofold: it differentiated between bees and wasps in
images, and it identified varroa mites on bees. This distinction is crucial for beekeepers
and environmentalists for effective bee population management and health monitoring.
For deployment, the model was served using TensorFlow Serving within a Docker container,
ensuring scalability and efficiency. The system architecture included a user-friendly
frontend for image uploads and a backend server hosted on Cloud Run. This server managed
request preprocessing and model prediction interactions.
Music Genre Classification with a Neural Network
Myself and two others built and trained a neural network that could classify songs into 10
different genres. As part of the project we tested a baseline which was trained on the basic
waveform amplitude of the audio file, and compared it to a model trained on 5 specifically
chosen features. Our baseline saw a 2x improvement over randomly guessing the genre and our
feature trained model saw a 6x improvement, with accuracy around 64% on
average.
The neural network used was a multi-layer feed forward network that was
built using Keras and TensorFlow. Data was pulled from the
GTZAN
dataset and was preprocessed into Excel using automated
Python
methods that we created. The entire project was written in a Jupyter Notebook
. More information about it is available
here.
Cooperative Autonomous System
The goal of this project was to have two autonomous vehicles - a land based robot and an air
based drone - cooperate together autonomously to complete a task. The task chosen for our
demonstration was garbage collection but the system could easily be adapted to other tasks
such as cleaning up oil spills.
The system implemented a Mask R-CNN neural network that detected litter in
images. The image processing and path planning was done in Python with
OpenCV and NumPY. Most hardware components and sensors
were programmed in C. The land based robot was powered by a
Linux
based RaspberryPi which also acted as an NGINX server for video streaming.
You can learn more about this project
here.
