Postural Classification by Image Embedding and Transfer Learning: An Example of Using the OWAS Method in Motor-Manual Work to Automate the Process and Save Resources

Forest operations often expose workers to physical risks, including posture-related disorders such as low back pain. The Ovako Working Posture Assessment System (OWAS) is widely used to assess postures in forest operations, but it requires expertise and significant resources. In this study, the use of image embedding and transfer learning was explored to automate OWAS classification. Over 5000 images from motor–manual cross-cutting operations were analyzed using two models: Google’s Inception V3 and SqueezeNet, both of which were integrated with neural networks via the Orange Visual Programming platform. The image vectors were fed into a locally run neural network (a multilayer perceptron with backpropagation) that was optimized for architecture and hyperparameters. The models were trained and tested using 20-fold cross-validation on the Posture and Action datasets, achieving accuracies of 84% and 89%, respectively, with Inception V3 outperforming SqueezeNet on both datasets. Predictions on unseen images yielded lower accuracies (50%–60%), highlighting the challenge of domain differences. These results demonstrate the potential of embedding-based transfer learning to automate postural classification with high accuracy, thereby reducing the need for expertise and resources. However, further research is needed to improve performance on unseen data and to explore alternative classifiers and embedding methods for better representation.