[5-1130-P] Other Categories (5th)

[5-1130-P-22] Data Extraction for Pig Weight Prediction Model

*Khin Dagon Win¹, Kikuhito Kawasue¹, Hsu Lai Wai¹, Kumiko Yoshida² (1. University of Miyazaki(Japan), 2. KOYO Plant Service(Japan))

Keywords:Weight estimation, Machine learning, 3D information, Random Forest, Multiple slits

Recently, automatic pig sorting systems have been popular to manage pigs in some pig farms. This systems automatically select pigs with appropriate weight for delivery. Normally, the pigs with over 115 kg are delivered in Japan. Therefore, this weight estimation system is essential to determine the maturity of pigs for shipment. A load cell is generally used in automatic sorting systems. However, it takes over 20 seconds to measure weight to detect stable weight. Sawdust is often used in pig house, but it can be attached to load cell and can lead to mechanic errors. Therefore, the use of load cell becomes big challenges to apply in actual pig farms.To overcome problems of load cell, we have developed an automatic pig weight measurement system using a camera. This system is composed on a camera, multiple slits and random dots projector. The camera with band-pass filter captures the pig image which enters into the system without influence on external luminous. Random dots and multiple slits are simultaneously projected to the pig body. Random dots projector is used to detect the location of pigs in the system and multiple slits projector is used to measure 3-dimensional shapes of pig body. Random dots projector is simultaneously projected to cover the whole surface of multiple slits. This measurement device is set up at the top of the system to detect back shape of pig body because the back shape can hold the definite growth conditions of pigs without being influenced by their daily nourishment levels. The image processing based weight estimation system consists of 3 steps: Extraction of pig from capture image, Quantitative analysis of the pig size from extracted image, Weight estimation from pig size using machine learning algorithm. Sawdust is often used in pig house. Moreover, those sawdust can be attached to a pig body. These attached sawdust can be influenced on extraction process of pig from captured images. In our system, Fast Fourier Transform (FFT) is applied to extract the pigs without being influenced by the surface situations of pig body. FFT detects the displacement of random dots to judge of existence of pigs in measurement area. 2-dimensional pig size information can be established with silhouette pig image. Furthermore, 3-dimensional pig size information is also considered to observe more specific growth conditions of pigs. For 3-dimensional information, it is needed to process slits image which are projected on pig body. Each slit location is detected to perform in the triangulations and 3D information such as length, girth and height are calculated. The adequate selection from 2D&3D information to estimate the pig weight is important and difficult process for our system. Therefore, Random Forest algorithm is utilized in our system. Random Forest randomly selects the samples from datasets and splits the data into several trees according to their features importance. The estimated weights are resulted by majority voting of its several trees. This method is adequate for pig weight estimation on practical conditions. The experimental results show the usefulness of our pig weight estimation system for automatic sorting system.