AGRICULTURAL RESEARCH COMMUNICATION CENTRE www.arccjournals.com/www.ijaronline.in *Corresponding author’s e-mail: mingxia@njau.edu.cn Indian J. Anim. Res., 52 (6) 2018 : 923-928 Print ISSN:0367-6722 / Online ISSN:0976-0555 Swine live weight estimation by adaptive neuro-fuzzy inference system Cedric Okinda, Longhen Liu, Guangyue Zhang and Mingxia Shen* College of Engineering, Nanjing Agricultural University, Jiangsu 210 031, P.R. China. Received: 01-10-2016 Accepted: 19-11-2016 DOI: 10.18805/ijar.v0iOF.7250 ABSTRACT Swine live weight is an important aspect in the production of pork products and also to Stockmen, with reference to market costs, feed conversion, and animal health. The objective of this study was to develop a contactless, stress-free method of swine live weight estimation by machine vision technology. This novel approach was based on image processing for features extraction and Adaptive Neuro-Fuzzy Inference System (ANFIS) for modelling. Firstly, the model determines which input combination holds the highest predictive ability, secondly, used the feature combination with the best predictive power to correlate to live-weight. The test results showed that the average relative error of our proposed system was about 3% and a standard deviation of 0.7%. Thus, development of a practical imaging system for swine live weight estimation by the proposed method is feasible. Key words: Adaptive Neuro-Fuzzy Inference System, Contactless, Features, Modelling, Predictive Power. INTRODUCTION Animal live weight information can be used to estimate the animal growth, feeds conversion efficiency, disease occurrences, body uniformity and market readiness (Wang, et al., 2008; Menesatti, et al., 2014; Li, et al., 2015; Wongsriworaphon, et al., 2015). Methods for assessing the live weight of swine are of great importance from several perspectives and have been widely studied and applied. Animal-predicated food products are customarily related to the ages of the animals that are utilized in their production, e.g., shrimp products (Pathumnakul, Piewthongngam et al., 2009) and pork products (Khamjan, et al., 2013; Plà- Aragonés and Rodríguez-Sánchez 2015). A 22 weeks old pig is ready for slaughtering for most pork products as the feed-to-live-weight ratio increases rapidly with age, thus, if a group of swine is not slaughtered in a particular week, the conversion of the feeds into weight will be less efficient in the following week. Therefore, matching appropriate size pigs to the feed products during slaughtering and victuals processing stages could amend the production efficiency by reducing the raw material procurement, inventory, shortage and perished costs (D’Souza, et al., 2004; Oliveira, et al., 2009; Apichottanakul, et al., 2012; Agostini, et al., 2013) One of the widely studied methods has been the body parts quantification approach (Topai and Macit 2004; Tasdemir, et al., 2011; TASDEMIR, et al., 2011; Tscharke and Banhazi 2013; Menesatti, et al., 2014). Self-accessed scales were developed to automated the animal’s front or hind leg mass estimation during feeding (Ramaekers, et al., 1995; Zaragoza 2009). However, it required extensive data processing to eliminating redundant data, hence prone to serious estimation errors (Williams, et al., 1996). Walk through scales have also been developed for dairy cows (Cveticanin and Wendl 2004; Alawneh 2011). However, accuracy depended on how the cows proceeded, whether they ambulated at a steady pace or not over the equipment at all. (Cveticanin 2003). Machine vision techniques have been applied to determine live weights and growth of various animals: (Minagawa and Murakami 2001) chicken (Mollah, et al., 2010), cattle (Tasdemir, et al., 2011; Menesatti, et al., 2014), swine (Pastorelli, et al., 2006; Wang; et al., 2008), and rabbit (Negretti, et al., 2010). Images of animals are mostly used to extract features which are used to correlate to the corresponding live weight, however, digital image-based methods for estimating an animal live weight require the animal be at an appropriate position and quite stationary otherwise the accuracy and computational speed would be compromised hence it is impractical in a real farm scenario (Kashiha, et al., 2014). Precedent studies sought to develop methodologies for quantifying a swine’s live weight while it was moving mundanely, in the feeding area, however, this mandated additional low-level illumination setups in the feeding area (White, et al., 2004; Wang, et al., 2008). Albeit the approaches mentioned above provided precise estimation, expeditious operation and were non-stressful to the swine, the quality of the digital images, the efficiency of the image processing technique and background determines the accuracy of the estimated live weight (Wongsriworaphon, et al., 2015).