Developing a Sow Vulva Volume Estimation Pipeline Based on LiDAR Imagery and Deep Learning

Abstract

Highlights An artificial intelligence-enabled imagery data processing pipeline. Sow vulva swelling can be quantified using imagery automatically. Biological signs of sow estrus are quantified. Abstract. Accurate and timely estrus detection is one of the critical factors for improving the conception rate of sows in swine reproduction. Estrus detection in the current industry relies on experienced breeding technicians using the back pressure test method, which is time-consuming, labor-intensive, and subjective. This study aimed to develop an image-processing pipeline that could predict the onset of estrus by monitoring the variation of the vulva volume of sows. A robotic imaging system consisting of a LiDAR camera was used to collect depth and infrared (IR) images of sows multiple times per hour. The developed image processing pipeline consisted of a deep learning model (MobileNet) to exclude images with defective postures (e.g., excreting posture, vulva region not showing) for vulva detection. The selected images were further processed to (1) segment sow vulva regions (3-dimensional surface) using a U-Net model, (2) exclude incorrectly segmented 3D vulva surfaces using a MobileNet model, and (3) calculate the vulva volume using the segmented 3D vulva surface. Results show that all three models achieved high test accuracy of 94.8% (MobileNet 1), 96.7% (U-Net), and 100% (MobileNet 2). The vulva volume was calculated using the developed pipeline for all the test sows. Daily average vulva volume showed a noticeable increase in all sows at 0-1 day prior to the onset of estrus. This study concluded that the developed image processing pipeline could automatically detect vulva swelling, which may be used as a biological sign indicating the onset of post-weaning estrus in sows. Keywords: 3D camera, Digital agriculture, Estrus detection, Robotic imaging system, Swine reproduction.