27August2026
14:00 Master's Defense Room 85 of IC2
Topic on
Comparative Analysis of Convolutional Neural Networks for Super-Resolution of Natural and Seismic Images
Student
Lucas de Magalhães Araujo
Advisor / Teacher
Edson Borin - Co-advisor: Sandra Eliza Fontes de Avila
Brief summary
Over the last decade, there has been a growing number of publications on Deep Learning applied to different problems in Geophysics. However, in most cases, the techniques are "borrowed" from another domain—Computer Vision models, for example, were developed in the context of images or videos, which have characteristics distinct from seismic images. This work proposes to explicitly explore the issue of technique transfer between domains. In particular, we will investigate the problem of Single Image Super-Resolution, comparing the relative performance of five Totally Convolutional Network architectures between the photographic and seismic domains. Specifically, the question guiding the research is whether the performance ranking between the different architectures will be maintained across the domains. To train and evaluate the models in the seismic domain, we constructed the Unicamp-NAMSS dataset, composed of 2.588 migrated 2D seismic images from 122 prospecting areas distributed across the American continent. The data were collected from the NAMSS platform, balanced, cleaned, and partitioned into training, validation, and testing partitions with geographic independence between them. In the photographic domain, we used the DIV2K dataset. Four experiments were performed: Super-Resolution with factors of 2× and 4× in both domains. The comparison between the models was made using the non-parametric Wilcoxon–Nemenyi–McDonald–Thompson method. We found no inversion in the partial ranking of performance between the domains: if one model was superior to another in the photographic domain, it was also superior in the seismic domain. The analysis of the results also revealed problems with the DIV2K dataset—not all images in the dataset are truly high-resolution—and with the NTIRE competitions—the DIV2K dataset is too small to distinguish the different techniques of the competition with statistical significance.
Examination Board
Headlines:
| Edson Borin | IC / UNICAMP |
| Alessandra Davólio Gomes | CEPETRO / UNICAMP |
| Alexandre Mello Ferreira | EEP/FUMEP |
Substitutes:
| Paula Sampaio Meirelles | CEPETRO / UNICAMP |
| Lucio Tunes dos Santos | IMECC / UNICAMP |