Chennakehava Akhil Pillalamarri

A segmentation-based computer vision pipeline for aviation image analysis. The project combines YOLO object detection with the Segment Anything Model (SAM) to detect and segment objects in aviation datasets including aircraft inspection images and airport operation scenes.

This project uses deep learning to automatically classify aerial disaster images into three categories — Intact, Major Damage, and Structural Failure. The images come from the LADI v2 dataset, which was collected by Civil Air Patrol pilots flying over US disaster zones between 2015 and 2023. Five different CNN models were built and trained from scratch to see which architecture performs best on this task. The main challenge was class imbalance — the dataset has far more Intact images than damaged ones, which makes it hard for the model to learn the minority classes. Several techniques were used to address this, including Focal Loss, balanced batch sampling, and data augmentation to increase the number of Major Damage training images. The DenseNet model performed best overall, while the Custom CNN Retrained was the only model that could detect Major Damage at a meaningful rate. GRAD-CAM heatmaps were generated to visually show which parts of each image the model focused on when making its predictions, helping explain the model's decisions in a human-readable way.