Goose Surface Temperature Monitoring System Based on Deep Learning Using Visible and Infrared Thermal Image Integration

Owing to increased biosecurity and industrial demands, the poultry houses in Taiwan are generally nonopen and closed types, with automatic environmental control and sensor equipment gradually being installed in such houses. Environmental sensors and poultry health monitoring systems are necessary to improve poultry feeding efficiency and safety. In this work, we developed a goose surface temperature monitoring system based on deep learning using visible image and integrated with infrared thermal image. This system could detect the geese in visible image and obtain the individual goose surface temperature automatically. This system consisted of an embedded system with the trained goose detection model, a visible camera, and an infrared thermal camera. The Mask R-convolutional neural network algorithm was employed to train the goose detection model by the collected goose images. The visible camera captured visible images in the poultry house, in which the geese could be identified by the trained goose detection model. The individual surface temperatures of the geese were obtained through integration of the visible and infrared thermal images. The developed monitoring systems were installed in the land and pool areas of a commercial goose house to monitor the surface temperature of the geese and achieved a precision of 97.1% and recall of 95.1%. In addition, the goose surface temperature of the pool area was observed to be lower than that of the land area. The collected individual goose surface temperature would be used as a management index to poultry house managers.

受生物安全要求提升与产业需求增长的双重影响，中国台湾地区的禽舍普遍采用非开放式封闭式结构，此类饲养舍正逐步加装自动环境控制装置与传感设备。环境传感器与家禽健康监测系统，是提升家禽饲养效率与饲养安全性的必备配套设施。本研究开发了一套基于深度学习（deep learning）的鹅只体表温度监测系统，该系统融合可见光图像（visible image）与红外热成像图像（infrared thermal image）技术。该系统可对可见光图像中的鹅只进行检测，并自动获取单只鹅只的体表温度。整套系统由搭载训练完成的鹅只检测模型的嵌入式系统、可见光相机与红外热成像相机组成。研究采用Mask R卷积神经网络（Mask R-convolutional neural network）算法，通过采集得到的鹅只图像训练鹅只检测模型。可见光相机可采集禽舍内的可见光图像，经训练完成的鹅只检测模型可识别图像中的鹅只。通过融合可见光图像与红外热成像图像，即可获取单只鹅只的体表温度。本研究将开发完成的监测系统部署于某商品肉鹅舍的陆地活动区与水池活动区，用于监测鹅只体表温度，最终实现了97.1%的精确率（precision）与95.1%的召回率（recall）。此外，研究观察到水池活动区内鹅只的体表温度低于陆地活动区。本次采集得到的单只鹅只体表温度数据，将作为管理指标供禽舍管理人员使用。