Color computer vision and artificial neural networks for the detection of defects in poultry eggs

A blood spot detection neural network was trained, tested, and evaluated entirely on eggs with blood spots and grade A eggs. The neural network could accurately distinguish between grade A eggs and blood spot eggs. However, when eggs with other defects were included in the sample, the accuracy of the neural network was reduced. The accuracy was also reduced when evaluating eggs from other poultry houses. To minimize these sensitivities, eggs with cracks and dirt stains were included in the training data as examples of eggs without blood spots. The training data also combined eggs from different sources. Similar inaccuracies were observed in neural networks for crack detection and dirt stain detection. New neural networks were developed for these defects using the method applied for the blood spot neural network development. The neural network model for blood spot detection had an average accuracy of 92.8%. The neural network model for dirt stained eggs had an average accuracy of 85.0%. The average accuracy of the crack detection neural network was 87.8%. These accuracy levels were sufficient to produce graded samples that would exceed the USDA requirements.