Comparison of Artificial Neural Network Training Algorithms for Predicting the Weight of Kurdi Sheep using Image Processing

Extended Abstract Introduction and Objective: Due to weakness, the occurrence of unwanted errors, the impact of the environment and exposure to natural events, human always make mistakes in their diagnoses of the environment or different topics, so that different people 's perception of a single and unique event may be very different and be diverse. Nowadays, with the development of image processing technology, human beings try to evaluate the speed and accuracy of their evaluation and diagnosis about objects, plants and animals by using hardware and software facilities and by using the features extracted from images related to objects, plants and animals. To increase and therefore has created a new technology called image processing and has developed it in various dimensions. Material and Methods: In order to identify the best artificial neural network training algorithm for estimating the weight of Kurdi sheep using digital image processing, lambs and adult animals at the sheep breeding station of North Khorasan province were weighed using a scale. During the weighing, some digital images were taken from the side view of sheep using a digital camera by discerning fixed distance. Image processing steps and feature extraction from images of sheep were done using GUI of MATLAB (R2010a) software. Then, three types of artificial neural networks were trained using different types of educational procedure, including Levenberg Marquarth (LM), Scaled Conjugate Gradient (SCG), and Bayesian Regularization (BR) Training algorithms. The extracted features from images were used as input and weight of sheep as output in the training steps of ANNs, and the accuracy of the ANN models in estimating the weight of sheep was compared. Results: As results, the accuracy of the trained ANNs with the three algorithms including SCG, BR and LM, in estimating the weight of sheep in the training phase was estimated to be 91.95, 94.74 and 94.94%, respectively. In the practical test, which was performed by presenting 20 images as a test to each ANN models, the trained ANNs with the SCG, BR and LM algorithm had 4.7%, 0.5% and 2.11% error in estimating the weight. The results showed that all three types of ANN training algorithms had acceptable accuracy to estimate the weight of sheep, meanwhile the accuracy of the artificial neural network trained with BR algorithm was better than the others. Conclusion: The performance of the proposed method based on image processing and the use of artificial neural network is accurate enough to estimate the weight of Kurdi sheep. It had better performance. Based on the results of the present study, it is quite possible to develop applications based on the use of artificial intelligence to weigh domestic animals, and use of this technology is recommended in several cases where there is no quick and easy access to the scales.