A Deep Metric Learning Approach for Content Based Image Retrieval in Rock Tomography: A Triplet Loss Study

• Anderson Silva UFMA
• Marcos Farias UFMA
• João Souza UFMA
• Daniel Pinto UFMA
• Filipe Belfort UFMA
• Mario Freitas UFMA
• Alexandre Araújo UFMA
• Alexandre Pessoa UFMA
• Aristófanes Silva UFMA
• Andrey Rodrigues PUC-Rio
• Marcelo Albuquerque PUC-Rio

Resumo

The automatic analysis of images in rock tomography datasets still is an unexplored task despite its high relevance in geological studies and characterization of materials. Traditional Content-Based Image Retrieval (CBIR) methods face difficulties when dealing with the structural complexity of these images while approaches based on deep learning still lack research specifically at this domain. In this paper we propose a CBIR approach using Deep Metric Learning with Triplet Loss function and different convolutional backbones. The results on the sandstone samples dataset has shown superior performance than the method found in the literature. The highlight was the DenseNet121 architecture, which obtained 99.23%±0.14 of F1-Score and 99.09%±0.16 of mAP@10. These results show the potential of the proposed approach to efficiently structure the similarity space in rock tomography images, demonstrating the relevance of the choice of a deep encoder when retrieving tomographic images in geological contexts.

Palavras-chave: Measurement, Graphics, Deep learning, Image retrieval, Decision making, Tomography, Rocks, Complexity theory, Faces