DenseDFFNet: Dense connected dual-stream feature fusion network for calf manure segmentation and diarrhea recognition

Abstract

Neonatal calf diarrhea is a globally prevalent disease, accounting for 57% of pre-weaning calf mortality. Early detection and intervention of diarrhea symptoms are critical for reducing morbidity and mortality while improving breeding efficiency. In intensive farming environments, it is challenging for staff to identify diarrhea symptoms in calves timely and effectively, as automated recognition methods for calf diarrhea remain underdeveloped. To address this issue, a non-contact calf diarrhea recognition method based on DenseDFFNet has been developed. By employing the multi-modal segmentation model Grounded-Segment-Anything (G-SAM) for manure segmentation, the difficulty of data annotation was significantly reduced and a fecal diarrhea segmentation accuracy of 96.45% was achieved in complex backgrounds. After segmentation, to mitigate abrupt pixel value transitions at image boundaries, a Parallel Convolutional Squeeze-and-Excitation (ParallelConvSE) module was designed, effectively integrating local and global features through parallel standard convolution and Squeeze-and-Excitation (SE) attention mechanisms. And the overall performance and generalization capability of the model was enhanced. For diarrhea classification, the DenseDFFNet module was introduced. In fecal classification tasks, the model achieved a test accuracy of 95.87%. When validated with video data, recognition accuracies for diarrhea and normal states reached 93.92% and 91.21%, respectively. Additionally, a self-propelled data collection system has been developed to enable efficient diarrhea recognition in complex commercial farming scenarios, offering a novel solution for calf health monitoring and early diagnosis. With its non-contact, efficient, and objective characteristics, the proposed method significantly reduces labor intensity and provides a robust technical solution for the recognition of calf diarrhea.