Detecção de faces e regiões de interesse em imagens de bovinos por meio de redes neurais convolucionais

Abstract

Livestock farming is an activity of great economic and social importance, and technology has played an increasingly important role in this sector. In this context, the present work is part of an intersection between Precision Livestock and Computer Vision. The motivation of this work arises from the need to identify the face of animals and its main regions of interest to solve problems such as the classification of facial expressions associated with pain or discomfort. From this perspective, it is proposed the application of models to detect face detection and its regions of interest in bovine images. To this end, one of the main steps of this work was the organization of a dataset annotated specifically for this purpose, containing images of cattle in different lighting conditions, angles and scenarios. This organization involved the manual collection, filtering and annotation of a large amount of images. The resulting dataset constitutes an important contribution of this work. Subsequently, studies were conducted on neural networks for training object detection models. Thus, two detection models were trained and evaluated from the neural networks SSD MobileNet V2 FPNLite 640 × 640 and YOLOv8. Regarding the evaluation criteria of the models, the experiments achieved satisfactory results given the problem addressed. Regarding the plotted graphs for each model, it was possible to observe that both networks had similar results to each other and consistent with the expectations of each graph. Finally, the qualitative analysis reinforced the results obtained with a visual study of the detections of both networks on representative examples of the test set. Overall, the results of the analysis demonstrated that both models are able to detect the face of cattle and their facial regions adequately, even under adverse conditions. In conclusion, this work demonstrates the potential of deep learning techniques for object detection in Precision Livestock. The results obtained are encouraging and suggest that these networks can be used to improve the efficiency and accuracy of management and production activities in rural areas. However, there are still challenges to be overcome, such as improving the robustness of the models and increasing the number of images in challenging situations in the dataset