Pig eye area temperature extraction algorithm based on registered images

Abstract

The body temperature of pigs is closely related to pig diseases. People have tried various methods to use infrared thermography technology for extracting pig body temperature. However, infrared images may vary with changes in pig body temperature. When the body temperature of pigs is abnormal, these body temperature extraction methods using only infrared images may not be able to locate the target area, thereby failing to complete the body temperature extraction operation. Taking the eye area as the target area, this study proposes apig eye area temperature extraction algorithmbased on registered images. This algorithm first aligns the infrared image with the visible image using registration techniques, then uses an object detection model to locate the pig's eye area on the visible light image, and finally completes the temperature extraction by combining the infrared temperature matrix. This algorithm uses visible images to locate the pig's eye area, fully utilizing the fact that visible images do not change with temperature, which can effectively avoid the problem that the pig's eye area may not be located by using only infrared images when the pig's body temperature is abnormal. To apply the pig eye detection model to pig farms and facilitate deployment on edge devices, this study conducts lightweight modification to the model by introducing the Ghost module to modify the neck and head structures of YOLOv7-tiny. The improved algorithm effectively reduces the number of parameters and computational cost. Compared with the original YOLOv7-tiny model, themean average precisiondecreased by 0.8%, while the number of parameters andcomputational costdecreased by 42.3% and 31.5%, respectively. The detection speed of the improved algorithm on the Jetson Nano (an edge device) increased by 12.8%. The improved algorithm achieves a balance betweendetection accuracyand detection speed, making it more suitable for deployment on edge devices.