Role: Machine Learning Engineer
Client: (R&D projects with Eastern Kentucky University)
Project Objective:
Primary objective of this project was to develop a computer vision system that utilizes drone imagery to generate valuable information and insights for pineapple farmers in Africa. By providing accurate statistics and maps, the system aims to help farmers make informed decisions, optimize their farming practices, and ultimately increase their yield and livelihood.
![Untitled](https://prod-files-secure.s3.us-west-2.amazonaws.com/b8baa53d-2b81-45ee-9cbc-adc4534b7afa/b00b5a2b-529e-418c-81d2-107229e7b40d/Untitled.png)
Main goal: Train multi-label object detection model using custom train data
Available labels (classes) for detected objects are like below:
0
- fruit
1
- flowers' to '1-flower
2
- without flowers' to '2-without flower
Key Achievements:
- Flowering Plant Identification:
- Developed a system that accurately identifies and localizes pineapple plants that have not flowered.
- This information is crucial for farmers, as it enables them to induce flowering in non-flowering plants using chemical or hormonal solutions, ensuring optimal fruit production.
- Fruit Counting:
- Implemented a feature that counts the number of pineapple fruits in the drone imagery.
- By knowing the expected production, farmers can make necessary investments and plan their resources effectively.
- User-Centric Design:
- Tailored the system to cater to the needs of various user types, including farmers, drone operators, and extension agents.
- Ensured that the information generated is easily accessible and understandable for all stakeholders.
- Image Preprocessing:
- Developed advanced preprocessing techniques to enhance the quality of drone images before applying machine learning algorithms.
- Implemented noise removal, sliding window resolution standardization, and geometric/color calibration to ensure accurate and reliable results.
- Multi-Label Object Detection:
- Developed a custom Convolutional Neural Network (CNN) model for multi-label object detection.
- Trained the model using a custom dataset, fine-tuned it for optimal performance, and deployed it for real-time analysis.
- The model effectively identifies and classifies pineapple fruits, flowers, and plants without flowers.
- Efficient Handling of Large Images:
- Developed a robust algorithm to process and analyze large-sized drone images efficiently.
- Implemented techniques such as object detection, object classification, sliding window, and feature prediction to extract meaningful information from the imagery.
Technical Stack: