融合轻量化网络与特征修复的鸡只健康状态在线监测方法

Online Chicken Health Monitoring Method Integrating Lightweight Networks and Feature Restoration

摘要:

针对笼养蛋鸡健康状态监测因鸡舍光照不足、鸡只行为多样性及笼架遮挡导致的鸡冠特征识别精度低、病鸡漏检率高等问题，本文提出一种融合多维度鸡冠特征的病鸡在线监测方法。首先，构建基于改进YOLO v8的鸡头方位识别模型，通过引入StarNet轻量化网络结构优化特征提取效率，结合SimSPPF模块增强多尺度特征融合能力，并嵌入GAM全局注意力机制强化关键区域识别，实现正位/侧位鸡头方位95.3%的精准判别。其次，在Lab色彩空间中对a*分量实施OTSU阈值分割，运用DeepFillv2图像修复网络修复因笼架遮挡导致的鸡冠区域缺损，构建完整的鸡冠形态特征图谱。通过侧位鸡冠面积、正位鸡冠宽度及颜色建立鸡只多特征融合健康决策模型，最终病鸡识别准确率为96.5%，漏检率约为2%。

Abstract:

Aiming to address the low accuracy of comb feature recognition and the high missed detection rate of diseased laying hens in caged poultry health monitoring, which are caused by insufficient illumination in poultry houses, diverse hen behaviors, and occlusion from cage structures, an online diseased-hen monitoring method was proposed based on multi-dimensional comb feature fusion. Firstly, an improved YOLO v8-based chicken head orientation recognition model was developed. By introducing the lightweight StarNet network structure to optimize feature extraction efficiency, integrating the SimSPPF module to enhance multi-scale feature fusion, and embedding the GAM global attention mechanism to strengthen key region recognition, the model achieved an orientation classification accuracy of 95.3% for frontal and lateral chicken heads. Secondly, OTSU threshold segmentation was performed on the a* component in the Lab color space, and the DeepFillv2 image inpainting network was employed to restore missing comb regions caused by cage occlusion, thereby constructing a complete comb morphological feature map. Based on lateral comb area, frontal comb width, and comb color, a multi-feature fusion health decision-making model was established. The final diseased-hen identification accuracy reached 96.5%, with a missed detection rate of 2%.