Vocábularios visuais aplicados à detecção de edifícios em fotografias históricas

Abstract

In this work, we propose a technique based on a bag-of-keypoints representation to identify images containing buildings in the photographic collection of Arquivo Público Mineiro (APM), the Archives of the State of Minas Gerais. We evaluate the proposed work using digitized versions of the original. Digitization is an effective procedure to make collections publicly available, while preserving them from deterioration, but it does not solve, by itself, all the problem of access, which are also related to allowing the user to locate the desired items among thousands. The application of techniques of Image Retrieval Based on Content (RIBC) could be very useful in describing and indexing of large historical collections, which requires considerable time and effort. However, Archive images tend to be more challenging to implement RIBC techniques than recent photos, due to loss of quality of the original source and several effects caused by degradation. The proposed technique is based on a visual vocabulary which is used to build a feature vector for an image. This representation has been proved robust to occlusion and variations due to pose, scale, illumination and several transformations. Results show that, despite of the poor quality of the images, the bag-of-keypoints representation is able to provide detection rates around 73% for images containing buildings, indicating the suitability of the proposed method for the task of detection.