Using deep learning to accurately detect sow vulva size in a group pen with a single camera.

This paper presents a non-contact method for the detection of changes in sow vulva size in a group pen. The traditional approach to estrus detection is manually pressing down on the back of the sow to elicit standing responses; however, this method causes undue distress for sows not in estrus. When a sow is in estrus, the vulva is red and swollen due to the presence of endocrine. Monitoring changes in vulva size to detect estrus with as little impact on the sow as possible is the focus of this study. This is achieved using a single camera combined with a deep learning framework. Our approach comprises two steps: vulva detection and vulva size conversion. Images of sows of Yorkshire, Landrace, and Duroc breeds were collected in group housing, and the vulva was detected through artificial markers and the network architecture of YOLO v4. Based on the internal and external parameters of the camera, the detected size was converted into millimeters and the results of manual measurement (MM) and automatic calculation combined to calculate the size of the vulva. Analysis of the calculated size compared with MM indicates that the object recognition rate of the system exceeds 97.06%, with a size error of only + 1.70 to -4.47 mm and high-calculation efficiency (>2.8 frames/s). Directions for future research include the automatic detection of pig width.

The size of a sow's vulva is an important indicator of sow estrus. Non-contact means of monitoring size changes for estrus timing would represent a significant contribution to the field of pig farming. This paper thus focuses on development of a system for the automatic detection of sow vulva size using a single camera combined with a deep learning framework. Experiments showed that the object recognition rate of the system exceeds 97.06%, the vulva size error is +1.70 to −4.47 mm, and the calculation efficiency is high (>2.8 frames/s).