Behavior Recognition Model for Pigs Based on Convolutional Neural Network (CNN) Model EfficientNet and Recurrent Neural Network (RNN) Model BiLSTM

This study proposes a video analysis method based on EfficientNet feature extraction and Bidirectional Long Short- Term Memory (BiLSTM) fusion for the precise identification of four behaviors (eating, activity, resting, and cleaning) of pigs in a pig farm. By fine-tuning the ImageNet pre-trained EfficientNet model through transfer learning to extract 1280-dimensional spatial features, and combining it with LSTM to capture temporal dependencies, the dynamic behavior recognition ability is significantly enhanced. Comparative experiments show that the BiLSTM model outperforms the LSTM model with an accuracy of 87.3% and a macro-average F1 score of 0.879 (+4.1% accuracy), particularly achieving a 6.0% F1 improvement in dynamic behaviors like cleaning and activity, and reducing the cross-misclassification rate by 28%-37%. This model, through bidirectional temporal modeling, can effectively analyze the spatiotemporal correlation during behavioral transition stages, providing a high-precision and lightweight solution for intelligent monitoring in animal husbandry. In the future, it can be further optimized for complex scene adaptation by combining with attention mechanisms.