教授
发布时间：2026-09-02
点击次数：
| 影响因子： | 10.3 |
| DOI码： | 10.1016/j.compag.2026.111954 |
| 所属单位： | 长江大学计算机科学学院 |
| 发表刊物： | Computers and Electronics in Agriculture |
| 刊物所在地： | Amsterdam, Netherlands |
| 关键字： | Group-housed pig behaviour; Liquid neural network; Object detection; Dynamic loss function; Precision livestock farming |
| 摘要： | The behavioural recognition of group-housed pigs constitutes a fundamental pillar for animal health monitoring and precision feeding within the realm of intelligent husbandry. However, environmental complexities—including fluctuating illumination, occlusion interference, and significant scale variations across the full growth cycle—severely constrain the generalisation capabilities of existing models. To address these challenges, this paper proposes PB-DETR, featuring a redesigned encoder and loss function within the framework. Inspired by time series, we innovatively propose an adaptive liquid time constant enhancement (ALTE) network in image data, integrate the liquid time constant (LTC) mechanism into the Transformer-based encoder, and introduce a dynamic enhancement module to exploit the temporal modelling capabilities of LTC. By projecting 2D image features into a characteristic evolution space and solving ordinary differential equations (ODEs) inspired by biologically motivated neural dynamics, this module simulates the continuous non-linear evolution of features along the spatial dimension, thereby effectively alleviating the model’s dependence on background information. Building upon this enhanced representation, a Nexus Pyramid Network (NPN) is further constructed to resolve scale inconsistencies across feature maps. By employing a two-stage iterative refinement strategy and focal feature units with dilated receptive fields, the NPN establishes bidirectional semantic interaction pathways across multi-scale features, ensuring robust and accurate spatial perception of pigs across varying sizes. With respect to the loss function, a Dynamic Quality Boost Loss (DQBL) is proposed, which adopts an exponential moving average (EMA)-based dynamic benchmark to evaluate the relative quality of predictions with respect to the model’s current performance, and adaptively adjusts prediction weights to guide continuous optimisation. To rigorously evaluate the reliability of the model’s generalisation, this study developed a large-scale dataset encompassing four typical husbandry environments. Beyond performance comparisons on standard validation sets, independent cross-dataset inference tests were conducted across four heterogeneous scenarios. The results demonstrate that PB-DETR outperforms state-of-the-art methods across key metrics, achieving an mAP of 0.893 while maintaining low computational overhead, with 14.58 M parameters and 49.12 GFLOPs. This validates its practical utility and potential for deployment in complex farming scenarios. |
| 合写作者： | Sisi Tong, Hongshen Guo, Bowen Tang, Jiayi Wang, Zhiliang Zhang, Wei Zhan |
| 第一作者： | Yuheng Guo |
| 论文类型： | 期刊论文 |
| 通讯作者： | Wei Zhan |
| 论文编号： | 20260830-001 |
| 文献类型： | 期刊 |
| 卷号： | 250 |
| 页面范围： | 111954 |
| 字数： | 11780 |
| ISSN号： | 0168-1699 |
| 发表时间： | 2026-05-31 |
