Leveraging museum specimens for AI-driven trait detection using computer vision: A case study on Zimbabwean fish

Natural history museums worldwide curate millions of preserved specimens and provide a vast untapped resource for scientific research. Advances in computer vision, particularly deep learning, now enable automated analysis of image data and several museums have made significant strides in digitising their collections. These efforts demonstrate the potential of digitised specimen images to support computer vision applications, including deep learning-based object detection and trait recognition. However, relatively few studies have explored the potential of preserved museum specimens as training data for automated trait detection. This study explores whether images of preserved fish specimens can be used to effectively train a deep-learning model, You Only Look Once (YOLO), to recognise morphological traits with key functional relevance in selected fish species from Zimbabwe. We also evaluated whether differences in model size (i.e. nano, small and medium) affected detection accuracy. Additionally, we evaluated these models' performance on fish images from outside the museum setting to assess their potential for broader applications. Our results showed that all model sizes achieved high detection performance on museum specimens, with mean recall exceeding 0.95 and mAP@0.5 values ranging from 0.75 to 0.78. The medium model showed slightly higher precision and stricter localisation (mAP@0.5:0.95) and the nano model slightly outperformed it in overall detection (mAP@0.5). However, performance differences between the nano, small and medium models remained small and were not statistically significant across repeated validation subsamples. This suggests that lightweight models can provide performance comparable to larger models for detecting traits, making them viable for museums with limited computational resources. When applied to external images from the Global Biodiversity Information Facility (GBIF), the models showed a noticeable decline in precision (0.073–0.166) and recall (0.153–0.251), highlighting the limited ability of models trained on a small set of curated museum specimens to generalise to real-world images. Overall, this study demonstrates that curated museum specimen images provide a structured and valuable data source for training computer vision models to extract morphological traits, support trait-based taxonomic interpretation and advance digital curation efforts. While the performance of the object detection models was robust under museum-specimen imaging conditions, limited transferability to GBIF images suggests that broader application will require expanded datasets that include images from additional institutions and data sources.