Abstract:
This project focused on the deployment of a deep learning model for species categorization and retraining on the cloud using the YOLOv5x architecture. The goal was to create a web application that allows users to upload images, classify them using the trained model, and continuously retrain the model with new data. The project also explored the use of MLOps best practices to enhance the project’s efficiency and scalability. The key challenges in deploying vision ML models on the cloud, such as data management, model training and retraining, and deployment infrastructure, were addressed. The YOLOv5x model’s ability to perform object detection, classification, and segmentation tasks, along with its use of cross-stage partial connections and anchor-based approach, made it suitable for the project’s requirements. The incorporation of MLOps best practices, such as continuous retraining and automated deployment pipelines, ensured that the model remains up to date with the latest data, resulting in improved accuracy and better decision-making. The project also looked at possibility of using AL techniques, since state of the art AL approaches typically rely on measures of visual diversity or prediction uncertainty, which are unable to effectively capture the variations in spatial context.