A Vision Machine for Detecting Fertile Eggs and Performance Evaluation of Neural Networks and Support Vector Machines in This Machine

In this research, a system is proposed for detecting fertility of eggs. The system is composed of two parts: hardware and software. The fabricated hardware provides a platform to obtain accurate images from inner side of the eggs, without harming their embryos. The software part includes a set of image processing and machine vision processes, which is able to detect the fertility of eggs from captured images, without any sensitivities to different types of eggs (e.g. with different thickness of the eggshell). In order to classify the fertile and infertile eggs, two classifiers based on Artificial Neural Networks (ANN) and Support Vector Machines (SVM) are designed and tested. It means that, to have a fully automatic fertility detection machine, we design two machine learning approaches using SVMs and ANNs to classify fertile and infertile eggs. That is, instead of using a predefined threshold values for distinguishing fertile pixels of egg images from infertile ones, we try to train the machine to do the job automatically. After training the machine using both classification algorithms, the performance of them are accurately investigated and measured in order to select the appropriate one. To evaluate the system, an egg image dataset is provided including 1200 images captured from incubated eggs. Extensive experiments are performed using the provided dataset, which confirm the reliable performance of the system. Comparisons with other fertility detection approaches applying different methods and algorithms confirm that the proposed machine outperforms more complex systems. Performance evaluations of the two proposed classifiers confirm that the SVM based classifier, with average detection accuracy of 50.57% at day 1 of incubation, 83.67% at day 2, 94.20% at day 3, 98.03% at day 4, and 98.91% at day 5, performs better than ANN based classifier, and it is also less sensitive against the reductions in training samples, which can be a serious issue when we are not able to provide more training samples.