Deep learning-based automatic dairy cow ocular surface temperature detection from thermal images

Abstract

Ocular surface temperature changes can be used as indices for evaluating the health statuses of dairy cows. However, the ocular surface temperatures of dairy cows in previous studies were mostly manually obtained from infrared thermal images captured by a thermal imager. Therefore, to automatically identify health disorders in dairy cows, we propose a lightweight and high-precision detection model based on the YOLOv4 framework, named GG-YOLOv4, which is used to automatically detect ocular surface temperatures from the thermal images of dairy cows. First, the CSPDarkNet module of the backbone YOLOv4 network with a large number of parameters and high complexity is replaced by a lightweight GhostNet module. Second, depth-wise separable convolution (DepC) modules are used to replace all 3 × 3 standard convolutions in the neck and head of YOLOv4 to greatly reduce the numbers of model parameters and calculations. Finally, based on the correlation between the positions of the eyes and the head, a global context network (GCNet) module is introduced to obtain perfect and effective feature information to compensate for the accuracy loss induced by the light part and improve the detection accuracy of the lightweight model. To evaluate the performance of the model, the thermal infrared images produced after data expansion are used as the test dataset. The experimental results show that the mean accuracy (mAP) of the GG-YOLOv4 network model is 96.88 %, the detection speed is 40.33f/s, and the model size is 44.7 M. Compared with other target detection models, the proposed method has better comprehensive performance in terms of detection accuracy, model size and detection speed. The ocular surface temperatures of dairy cows are obtained by this detection model and compared with the ocular surface temperatures manually extracted by FLIR-Tools. The comparison results show that the average absolute temperature extraction errors in the left and right eyes are 0.051 °C and 0.042 °C, respectively, and the average relative temperature extraction errors in the left and right eyes are 0.14 % and 0.11 %, respectively. The developed method realizes the rapid and accurate positioning of the key parts of dairy cows and can be used for high-precision ocular surface temperature extraction, providing a basis for the automatic diagnosis of dairy cows’ health statuses.