Feedback System for Improving Capturing Quality and Quantity of Livestock Images Using Deep Learning Technology

Abstract

Livestock body parameters like shape, horn, teeth, muzzle, and udder provide useful information to determine livestock age and health. It is very difficult to continuously monitor and measure these parameters for 300 million bovine animals in India. We developed a Deep Learning (DL) based intelligent Livestock Health Monitoring System (LHMS) which derives these parameters from the livestock images. We developed a mobile application for Veterinarians and livestock Artificial Insemination Technicians (AIT) to collect and monitor livestock data and images throughout their pregnancy lifecycle. Though AIT captured 1.87 Lakh livestock data since 2016, it had only 1000 images. We conducted multiple iteration of the Design Thinking (DT) research to understand the challenges in the image capturing process. It was difficult for a human to see each image and provide feedback to the AITs about quality of images. DL models revealed the poor quality of the images, such as missing livestock as well as noisy and blurred images. Model accuracy decreased due to this. To address this challenge DL were methods to analyze the image, train system and generated an AIT Image Score (AIS) based on factors like quantity of images, accuracy of images, frequency of upload, geo-location etc. Based on AIS, we created a personalized feedback message and training instructions on how to click and collect images for each AIT. This paper captures our experiences on use of DT approach, which resulted in an 80% jump in image quantity over a three month study period and 78% improvement in the quality of the images.