Enhancing DNN-based Damage Detection and Flood Segmentation in Complex Outdoor Environments

Deep neural networks (DNNs) have achieved significant advancements in computer vision tasks such as object detection and semantic segmentation. However, their deployment in complex real-world environments, such as industrial inspection and disaster response, remains challenging due to background clutter, occlusions, and computational constraints. This thesis introduces Region Augmented Knowledge Distillation (RAKD), a novel training framework designed to improve the performance of state-of-the-art (SOTA) DNN models in object detection and flood segmentation. RAKD leverages a Teacher-Student distillation paradigm in which the Teacher model is trained on augmented images that emphasize foreground regions, enabling it to develop more discriminative feature representations. During training, the Student model learns from the Teacher’s feature maps while processing real-world images, improving robustness against background noise and environmental variability. Extensive experiments demonstrate the framework’s effectiveness in two critical applications: damage detection in industrial pipelines and flood segmentation from UAV imagery. The proposed method consistently outperforms baseline training and traditional feature distillation approaches across multiple SOTA models. Ablation studies confirm that the choice of augmentation strategies significantly influences knowledge transfer. For instance, Remove Background and Sharpen Foreground preprocessing is optimal for industrial damage detection, whereas Blurred Background and Equalized Foreground preprocessing enhances floodsegmentation performance. Additionally, the selection of appropriate loss functions, such as Cosine Similarity loss for object detection and Mean Squared Error (MSE) loss for flood segmentation, results in substantial improvements in precision and recall. The experimental results indicate that RAKD consistently surpasses baseline methods across multiple evaluation metrics. Specifically, in damage detection, it improves the mean Average Precision (mAP) of the YOLOv6s baseline by 0.031 (+12%), while increasing mAP 50 by 0.045 (+9%). In flood segmentation, RAKD enhances the performance of PIDNet, achieving an improvement of 0.02 (+2.5%) in mean Intersection over Union (mIoU), the highest observed performance gain among tested models. Additionally, the application of optimized knowledge distillation strategies significantly reduces false positives and false negatives, leading to more reliable predictions in real-world scenarios. RAKD offers a scalable, architecture-agnostic, and computationally efficient solution for real-world vision applications. Its ability to generalize across tasks highlights its potential for broader applications, including autonomous inspection, disaster management, and environmental monitoring. The results of this study provide a foundation for future research in adaptive knowledge distillation techniques, particularly for deployment in dynamic and resource-constrained environments.