Binning of biological image based on unsupervised semantic segmentation and no-reference image quality assessment

Accurate and efficient species classification is crucial for various applications, and machine-learning approaches have been widely employed for image-based taxonomic identification. However, challenges arise due to the variability in image quality and the need for extensive dataset preparation. In this study, we propose a model that combines unsupervised semantic segmentation by distilling feature correspondences (STEGO) and a no-reference image quality assessment model to assess the quality of images for taxonomic identification. STEGO enables the counting of objects in an image, while the image quality assessment model evaluates subject articulation. The model is applied to a dataset of Bombus hypnorum images from the Global Biodiversity Information Facility (GBIF) after data cleaning. The effectiveness of the model is demonstrated through the segmentation and scoring of images, providing a valuable tool for image recognition and quality assessment in taxonomic studies.