Content-Based Image Retrieval for Fine-Grained Aircraft Using Deep Learning and Cosine Similarity

Traditional keyword-based retrieval is based on textual metadata, making keyword processing time-consuming, biased, and manual. With the increase in digital images, this system becomes ineffective. To solve this, Content–Based Image Retrieval (CBIR) offers a computerized and more accurate alternative. In this study, we focus on fine-grained classification that distinguishes subcategories within a broader category, such as Boeing 777 and Boeing 737. This is crucial for understanding subtle visual differences. Traditional image search methods for finding specific aircraft in large databases depend on keywords and metadata, which can be inconsistent, incomplete, or misleading. For example, one user might label an aircraft as a twin-engine jet, while another describes it as a commercial aircraft. If the database does not consist of exact terminology or expressions, the system may fail to retrieve visually similar matches. These limitations indicate that existing image retrieval systems are limited in analyzing the actual visual content of images. CBIR replaces keywords with visual search, providing more accurate results—especially when dealing with complex or fine-grained. Our approach involves feature extraction from VGG16’s GAP layer, combined with XGBoost and Cosine similarity, applied to the FGVC-Aircraft dataset, which contains 10,000 images across 100 classes. The model achieves an accuracy of 90.1%, with a Macro-F1 score of 83%, and the top retrieval results show strong visual consistency across aircraft models. Applications include airport security and surveillance, defense and military, the aviation industry, public safety, and commercial sectors. Future improvements could involve feature fusion and advanced similarity metrics.