Tens of millions of images from biological collections have become available online over the last two decades. In parallel, there has been a dramatic increase in the capabilities of image analysis technologies, especially those involving machine learning and computer vision. While image analysis has become mainstream in consumer applications, it is still used only on an artisanal basis in the biological collections community, largely because the image corpora are dispersed. Yet, there is massive untapped potential for novel applications and research if images of collection objects could be made accessible in a single corpus. In this paper, we make the case for infrastructure that could support image analysis of collection objects. We show that such infrastructure is entirely feasible and well worth investing in.

Purpose: Historical newspaper collections provide a wealth of information about the past. Although the digitization of these collections significantly improves their accessibility, a large portion of digitized historical newspaper collections, such as those of KBR, the Royal Library of Belgium, are not yet searchable at article-level. However, recent developments in AI-based research methods, such as document layout analysis, have the potential for further enriching the metadata to improve the searchability of these historical newspaper collections. This paper aims to discuss the aforementioned issue.Design/methodology/approach: In this paper, the authors explore how existing computer vision and machine learning approaches can be used to improve access to digitized historical newspapers. To do this, the authors propose a workflow, using computer vision and machine learning approaches to (1) provide article-level access to digitized historical newspaper collections using document layout analysis, (2) extract specific types of articles (e.g. feuilletons – literary supplements from Le Peuple from 1938), (3) conduct image similarity analysis using (un)supervised classification methods and (4) perform named entity recognition (NER) to link the extracted information to open data.Findings: The results show that the proposed workflow improves the accessibility and searchability of digitized historical newspapers, and also contributes to the building of corpora for digital humanities research. The AI-based methods enable automatic extraction of feuilletons, clustering of similar images and dynamic linking of related articles.Originality/value: The proposed workflow enables automatic extraction of articles, including detection of a specific type of article, such as a feuilleton or literary supplement. This is particularly valuable for humanities researchers as it improves the searchability of these collections and enables corpora to be built around specific themes. Article-level access to, and improved searchability of, KBR's digitized newspapers are demonstrated through the online tool (https://tw06v072.ugent.be/kbr/).

In recent years, many scientific institutions have digitized their collections, which often include a large variety of topographic raster maps. These raster maps provide accurate (historical) geographical information but cannot be integrated directly into a geographical information system (GIS) due to a lack of metadata. Additionally, the text labels on the map are usually not annotated, making it inefficient to query for specific toponyms. Manually georeferencing and annotating the text labels on these maps is not cost-effective for large collections. This work presents a fully automated georeferencing approach based on text recognition and geocoding pipeline. After recognizing the text on the maps, publicly available geocoders were used to determine a region of interest. The approach was validated on a collection of historical and contemporary topographic maps. We show that this approach can geolocate the topographic maps fairly accurately, resulting in an average georeferencing error of only 316 m (1.67%) and 287 m (0.90%) for 16 historical maps and 9 contemporary maps spanning 19 km and 32 km, respectively (scale 1:25,000 and 1:50,000). Furthermore, this approach allows the maps to be queried based on the recognized visible text and found toponyms, which further improves the accessibility and quality of the collection.

Since herbarium specimens are increasingly becoming digitized and accessible in online repositories, an important need has emerged to develop automated tools to process and enrich these collections to facilitate better access to the preserved archives. Particularly, automatic enrichment of multi-specimen herbaria sheets poses unique challenges and problems that have not been adequately addressed. The complexity of localization of species in a page increases exponentially when multiple specimens are present in the same page. This already challenges the performance of models that work accurately with single specimens. Therefore, in this work, we have performed experiments to identify the models that perform well for the plant specimen localization problem. The major bottleneck for performing such experiments was the lack of labeled data. We also address this problem by proposing tools and algorithms to semi-automatically generate annotations for herbarium images. Based on our experiments, segmentation models perform much better than detection models for the task of plant localization. Our binary segmentation model can accurately extract specimens from the background and achieves an F1 score of 0.977. The ablation experiments for multi-specimen instance segmentation show that our proposed augmentation method provides a 38% increase in performance (0.51 mAP@0.9 versus 0.37) on a dataset of 1,500 plant instances.

The digitization of image archives across the globe has opened up vast collections of libraries, museums, and cultural heritage institutions. These collections provide valuable historical information to the public and researchers. Many image collections have little metadata describing who or what is depicted in a structured format, making it difficult to search for specific persons. This work presents a facial recognition pipeline to enrich these collections by recognizing the persons in each image. A reference dataset of over 6,000 known persons was constructed and facial recognition was performed on a dataset of over 150 thousand images. Detected faces were matched with the known faces using a similarity score on the face embeddings. We developed an interactive labeling tool to efficiently validate the face recognition predictions. A total of 182 thousand detected faces were labeled with this tool. Using a minimum similarity score of 0.5, the face recognition model achieved a precision of 0.936 and identified over 62 thousand persons from the image archives. We show how clustering can be used to identify new persons that were not included in the reference dataset. Furthermore, we highlight the potential of facial recognition to enhance the accessibility of the collections and offer new insights.