Article Preview
TopIntroduction
Most animal species show their diversity in the form of breeds. It is necessary to identify these breeds for various reasons including implementation of the Global Plan of Action for Animal Genetic Resources (AnGR) as framed by Food and Agricultural Organization of the United Nation. Phenotypic and genotypic characterizations are the two pillars for recognition of breeds of domestic animals (FAO, Molecular Genetic Characterization of Animal Genetic Resources, 2011) and (FAO, Phenotypic Characterization of Animal Genetic Resources, 2012). The phenotypic characterization encompasses identification, quantitative and qualitative description, documentation of populations, information of the natural habitats and production systems of a breed. The genetic-characterization is done to understand the diversity and distinctiveness of a genetic resource for the purpose of framing policy for improving the value of the resource and its scope for utilization.
India is one of the few countries in the world, which has a major contribution to the international livestock-gene-pool and improvement of animal production in the world. It possesses about 132.74 million (FAOSTAT, 2018) goat population which is the second highest in the world after China. The country has so far registered 34 goat breeds as per report (NBAGR, 2015) of the ICAR-NBAGR, Karnal, Haryana; the premier institute, engaged in identification and characterization of livestock and poultry genetic resources of India. Declaration of breed identity by phenotypic characters seems unsatisfactory many times because of variations within a breed as well as presence of similar looking breeds especially, when purity of the breed is concerned. Genotypic characterization involves availability of expensive laboratory facilities. As a result, a new method for breed characterization is needed that should be accurate, economically viable and can be easily availed by the common people.
The aim of this paper is to build a goat breed dataset that can serve a variety of different purposes including research, training, planning, public awareness, decision-making and many more and propose a new approach to distinguish goat breeds with the help of images through deep learning-based object recognition technology. Since images are prone to noise and distortion like low quality images, low light images, images where the animal is partially obstructed, so this approach should also be resistant to such anomalies and provide consistent result for all conditions. This process would complement as well as supplement the existing system of breed identification technique non-invasively and also would be able to eliminate ambiguity in selecting pure breed animals.
In this paper, a novel goat breed dataset is constructed by capturing almost two thousand images of pure-breed goats which have been reared and maintained by government research organizations in India. The images are captured randomly with the help of a mobile device (Samsung A6) in such a way that the full goat is visible without imposing stress to animals. 85% of the captured images are used for training and the rest for testing. Since any deep learning model requires a large number of training images for getting good accuracy, the number of training images are increased by applying geometric transformations such as rotation, translation and scaling to the original images (augmentation). For object localization during test time, each training image needs to have coordinate information about the exact location of the object inside the image which is used by the bounding box predictor algorithm as the ground truth data. This is done using an image annotation tool by drawing a bounding box around the goat in the training images. Faster R-CNN with Inception-ResNet-v2 as the feature extraction backbone, trained on the COCO dataset is used as the pre-trained model. Then this model is fine-tuned by only re-training the fully connected layers with the help of the training set images for 100k iterations. After training is completed the model is tested on the test set of images and results are obtained. No augmentation and annotation are done on the test set images. In the output, each test image consists of a single bounding box around the goat describing its breed and prediction accuracy (Figure 1). During testing, the prediction accuracy is set manually, that means in any test image if the model predicts lower accuracy than the prescribed value no bounding box is drawn.
Figure 1. The Process of Goat Breed Identification and Localization