Título: Robust Landmark Recognition with Out -of-Distribution Detection using Deep Features and K -Nearest Neighbors
Autores: Lucas Ferreira Pereira, Pedro Herique Nascimento Castro, Valéria Santos, André Luiz Carvalho Ottoni, Gladston Juliano Prates Moreira & Eduardo J. S. Luz
Resumo: Recognizing landmarks from diverse, real-world images is challenging due to variations in viewpoint, illumination, and occlusion. A critical, often overlooked, aspect is the ability to reject images that do not depict any known landmark (Out-of-Distribution, OOD, samples). This paper proposes a robust system for landmark recognition that integrates a deep convolutional neural network (CNN) for feature extraction and classification with a k-Nearest Neighbors (KNN) based approach for OOD detection. We leverage transfer learning with a DenseNet-201 architecture, fine-tuned on a diverse landmark dataset. The proposed system achieves 97.5% classification accuracy on In-Distribution (ID) landmark images on the Visual Chine Dataset. Our KNN-based OOD detection method, using features from the trained DenseNet, achieves a 97.0% True Positive Rate for ID samples at a 95.0% recall threshold (TPR@95TPR), effectively measuring precision at high ID recall against OOD samples, demonstrating its efficacy in distinguishing landmarks from irrelevant scenes. This combined approach offers a practical solution for real-world tourist applications.
Palavras-chave: Landmark Recognition; Deep Learning; Convolutional Neural Networks; Out-of-Distribution Detection; k-Nearest Neighbors; Transfer Learning; Tourism Applications.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175623
Artigo em PDF: CBIC_2025_paper1175623.pdf
Arquivo BibTeX:
CBIC_2025_1175623.bib