Answers are grounded in this paper and cite their sources.
Forest workers are at risk of developing musculoskeletal disorders, and several methods have been employed to evaluate these risks. The Ovako Working Posture Analysing System (OWAS) is popular in the industry but can be resource-intensive when used conventionally. In this study, we developed a comprehensive annotated image dataset containing 23,000 images, drawn from a wide range of operations under diverse conditions. The dataset was augmented through translation and reflection techniques, and it was randomly divided to allocate 70%, 15%, and 15% of the data for training, validation, and testing, respectively. Subsequently, we employed ResNet-50, MobileNet-v2, ShuffleNet, and GoogLeNet models to evaluate classification performance through a comparative approach. This involved using stochastic gradient descent with momentum, a batch size of 10, 6 epochs, and a learning rate of 0.0001. Classification performance was assessed using accuracy and F1-score to identify the best-performing model for fine-tuning, which was conducted in two steps following the same data division strategy. The best-performing model was first tuned for learning rate (0.001 and 0.00001) and then for batch size (32, 64, and 128). Decisions regarding the optimal combinations of learning rate and batch size were based on accuracy and F1-score; the model was then retrained, validated, and tested using the updated parameters. Following the initial comparison experiment, results indicated that all models achieved high accuracy, with testing accuracies of 89.19%, 89.91%, 90.90%, and 91.22% for ShuffleNet, MobileNet-v2, GoogLeNet, and ResNet-50, respectively. The corresponding F1-scores for these models were 79.86%, 79.95%, 70.96%, and 77.92%. After fine-tuning ResNet-50, we found that a learning rate of 0.001 coupled with a batch size of 32 yielded improved performance, with classification accuracy and F1-score of 96.34% and 91.47%, respectively. Using these models offline can provide numerous benefits, including automation, reliability, speed, and accuracy, facilitating the evaluation of large datasets, which is often limited in conventional assessments. Furthermore, deploying the models in an online approach will offer real-time feedback on working postures and aid in early intervention measures, thereby contributing to improved health and safety in forest operations.
No takes yet. Share an insight, caveat, or question.
Forkuo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: