A Real-Time Drone Image Processing System using Deep Learning and Big Data Technology

In this research, we delve into advanced image segmentation techniques applied to drone imagery for various environmental and surveillance applications. By leveraging state-of-the-art models such as UNet, DeepLabV3, Manet, and Feature Pyramid Network (FPN), our goal is to achieve high precision in segmenting complex aerial scenes. Each of these models possesses unique strengths and weaknesses; hence, we employ an ensemble technique, weighted averaging, to harness their combined capabilities for superior results. Additionally, we incorporate image augmentation techniques to simulate various weather conditions such as haze and raindrops, enhancing the robustness of our models. To manage real-time data efficiently, we implement a streaming pipeline using Apache Kafka and Apache Spark, ensuring scalable and effective processing. Our methods demonstrate significant performance improvements when trained on the original dataset and the combination of original dataset and augmented dataset compared to conventional methods.