DLRG@ NakbaArchiveClassifier Shared Task: Deep Transfer Learning for Destruction Detection in Nakba Archive Images Using EfficientNet-B3

Abstract

Automatic identification of destruction in conflict-affected regions is an important task for humanitarian monitoring and historical documentation. Visual analysis of destruction scenes can assist researchers and policy makers in understanding the extent of damage in affected areas. This paper presents a deep learning-based image classification approach for identifying destruction and non-destruction scenes in Nakba-related images. The problem is formulated as binary image classification on Nakba images. A transfer learning approach using EfficientNet-B3 is adopted to learn discriminative visual features from Nakba images. Experimental evaluation shows that the proposed model achieved an Weighted F1-score of 83.87 % and an overall classification accuracy of 85.57 % and secured 10th rank in the competition. The results demonstrate that our proposed pre-trained method can effectively capture structural damage patterns and visual cues associated with destruction scenes. Code: https://github.com/kannanrrk/NakbaImageClassifier

Resources

Details