Abstract
Accurate recognition of diarrhea-related behaviors in calves is crucial for automated health monitoring in smart farming. However, existing methods face challenges including insufficient temporal continuity and limited capability in capturing periodic or abrupt behavioral changes. To address these issues, this study proposes a lightweight tri-branch time-frequency fusion network named self-distillation with no labels version 2 (DINOv2)–Transformer ODE (ODE)–Fourier-Wavelet (DTOFW) for calf diarrhea behavior recognition. The network employs DINOv2 as a frozen backbone for high-level semantic feature extraction and integrates three complementary branches for comprehensive behavioral modeling. The Transformer-ODE branch models continuous temporal dynamics to compensate for information gaps caused by discrete sampling. The Fourier attention branch adaptively emphasizes global periodic patterns in the frequency domain. The Wavelet attention branch captures local transient and non-stationary abnormalities associated with diarrhea. Experimental results on the Jinnan calf behavior dataset show that DTOFW achieves 95.32% recognition accuracy with 1.33 million trainable parameters in the tri-branch fusion head. Compared with the tested baseline models, DTOFW improves accuracy by up to 8.71 percentage points while maintaining a low number of trainable parameters in the fusion head. The proposed model provides an effective and efficient solution for fine-grained abnormal behavior recognition in intelligent livestock farming.