Avaliação do escore de condição corporal por imagens digitais em fêmeas suínas gestantes

Abstract

Monitoring body condition is one of the most critical practices in the production management and nutritional monitoring of female swine, performed at strategic moments to improve the reproductive performance rates of the herd population. However, these visual evaluations and/or caliper measurements are prone to inconsistencies due to personal perceptions of the ideal phenotype. Continuous evaluation of phenotypic characteristics under morphological assessment allows for nutritional adjustments that optimize sow productivity. This process, which requires time, labor, and involves stressful manual interventions for the animals, can be automated through the collection of 2D digital images of the animals. To minimize the bias of subjective BCS assessment, this study developed and evaluated a deep learning-based approach using pre-trained neural networks for automated classification in pregnant sows and gilts. Aiming to identify the architecture that provides the ideal balance between generalization capability and precision, 190 PIC® Camborough® females were used, with 760 2D digital images captured using a Samsung Galaxy A34 smartphone, considering 4 collections from 190 females during the gestation period, as follows: 567 images for the Ideal class, 140 images for the Under class and 53 images for the Over class. The grid search technique was used to execute 60 experiments, varying the hyperparameters, in addition to obtaining globally optimal combinations for the machine hyperparameters through the adjustment of the PSO algorithm. Furthermore, 80% of the data were used for training, within this set, 20% were used for validation (467 images for training and 116 images for validation) and subsequently, 20% of the dataset (177 images) were used for testing. All convolutional neural networks and hybrid architectures were evaluated using the weighted global F1-Score, the F1-Score per class, and accuracy on the test dataset, as well as the loss function. After the experimental configurations and the correction adjustment between hyperparameters, performance metric results for the convolutional neural networks were obtained on the test set. The experiment for the ResNet50 architecture demonstrated the greatest robustness and consistency, achieving sensitivity for both minority classes, with an accuracy of 70% and an F1-Score per class of 25.4%, 23.5% and 81.5% for Under, Over, and Ideal, respectively, in addition to a weighted F1-Score of 66.5%. The experiments conducted in this study demonstrated that the direct application of transfer learning to body condition in females presents challenges that go beyond the choice of neural network architecture, such as dataset quality. It was further observed that, despite hyperparameter optimization and the application of class weights, the final model encountered difficulties in discriminating against minority classes for all hybrid architectures. The results of this study suggest that the primary limitation lies in the feature extraction stage, likely due to the visual anatomical similarity between the BCS classes.