Título: Deep learning algorithm for human body type assessment from digital images
Autores: Antonio Ricardo A. Brasil, Fabian Tadeu do Amaral, Licia C. S. de Lima Oliveira, Girlandia A. Brasil, Klaus Fabian Cocô, Patrick Marques Ciarelli
Resumo: This paper presents a novel convolutional neural network (CNN) based regression approach for automated somatotype estimation from digital images. Unlike previous studies that classify individuals into predefined somatotype categories or rely on manual anthropometric measurements, our method directly estimates continuous somatotype component scores, leveraging advanced image segmentation (PointRend) to represent body shapes from digital images accurately. The innovation lies in integrating a segmentation technique with regression-based CNNs, offering higher accuracy and accessibility compared to prior automated methods. We applied advanced image segmentation (PointRend) to a dataset of 46 human body images and trained CNN models to directly regress endomorphy, ectomorphy, and mesomorphy values. Hyperparameter optimization was performed systematically to ensure optimal performance. Our CNN-based regression approach achieved comparable accuracy to traditional anthropometric methods (MSE < 0.16) while eliminating manual intervention. The best result for endomorph estimation (MSE = 0.15) surpasses previous image-based preprocessing methods (MSE = 0.44). The proposed method demonstrates significant potential for automated somatotype estimation, providing an accurate, accessible alternative to traditional measurement methods. This innovation facilitates wider applications in health, sports, and biomedical engineering contexts.
Palavras-chave: convolutional neural network; deep learning; somatotype; regression.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1190378
Artigo em PDF: CBIC_2025_paper1190378.pdf
Arquivo BibTeX:
CBIC_2025_1190378.bib