Treinamento e avaliação de redes neurais convolucionais (CNNs) para a identificação de comportamento canino em imagens

Abstract

Automatic action recognition can profoundly transform the way we interact with and care for pets. When applied to monitoring canine health, this technology has the potential to detect early signs of disease or discomfort, allowing for quick and effective interventions. This work aims to develop and validate a Convolutional Neural Network (CNN) model capable of classifying, from a single image, whether a dog is scratching or not. The scratching action, when performed excessively, can be a sign of food or environmental allergies, fleas, ticks, or other parasites, as well as skin conditions such as dermatitis and infections. The CNN-DATAUG-RELU-L2-DROPOUT model was designed with an architecture that utilizes convolutional layers with ReLU activation functions, L2 regularization, and dropout. The model was trained and evaluated on a specific dataset, achieving an accuracy of 87.83%, a precision of 90.23%, and a cobertura of 89.31%. These results indicate good performance under controlled conditions. Although action recognition models often use sequences of frames, this study demonstrates that analyzing individual frames can be a simple and effective approach.