Título: Deep Learning -Based Metadata Inference for Photographs Captured with Manual Lenses: Challenges and Opportunities
Autores: Luis Alfredo da Silva, Rafael Stubs Parpinelli
Resumo: This paper presents a deep learning approach to infer missing metadata, specifically focal length, aperture, and subject distance, from photographs taken with manual lenses. The lack of lens metadata on photographs taken with those lenses (due to their inability to electronically communicate metadata) poses challenges for photographers, archivists, and researchers who rely on this information for image organization, forensic analysis, and computational photography applications. Motivated by this, we develop a ResNet-based model adapted for regression tasks and train it on a diverse dataset of 2,200 original images, augmented to approximately 28,000 usable, representative samples. The model demonstrates encouraging results in both aperture and focal length estimation (MAPE < 26%).
Palavras-chave: Metadata inference; Deep learning; Convolutional Neural Networks; Dataset augmentation; Computational photography.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1139268
Artigo em PDF: CBIC_2025_paper1139268.pdf
Arquivo BibTeX:
CBIC_2025_1139268.bib