| The processing of poultry eggs for human consumption has four major steps--collecting, washing, grading, and packaging. The collecting, washing, and packaging steps have been mechanized. However, the egg grading step, in which eggs are inspected for defects such as blood spots, cracks, and dirt stains, is still done manually. Typically, a grader must inspect a dozen eggs per second and make decisions on whether to allow an egg to pass, reject and remove it, or send it to be rewashed. This leads to overpull, where good eggs are graded as defective, and underpull where defective eggs are undetected. Automation of the egg grading process is desirable since it promises to help control costs, reduce the work load on graders, and improve the quality control process.; Neural network models were developed to identify eggs with defects using gray scale images. A gray scale computer vision system was used to obtain images of grade A eggs and eggs with a single type of defect. Image histograms based on the intensity level were constructed. For each type of egg defect, a neural network model was developed using the histograms of eggs with the defect and eggs without that defect. The neural networks were tested and validated on independent data sets. Accuracies of 85.6%, 90.0%, and 80.0% were achieved by the blood spot, crack, and dirt stain detection neural networks, respectively. The blood spot and crack detection neural networks were able to produce graded samples that would exceed the USDA's requirements. The dirt stain neural network was not able to meet the USDA's specifications.; Other neural networks were developed using color images of eggs. A similar approach was used in developing neural networks with a color computer vision system as with the gray scale system. The use of a color computer vision system improved the accuracy of the neural networks. The accuracies were 92.8%, 87.8% and 85.0%, for blood spots, cracks, and dirt stains, respectively. These accuracy levels were sufficient to produce graded samples that would pass USDA inspections.; An expert system was developed to sort eggs into use-based categories. The expert system used the outputs of the neural networks to make sorting decisions. Variable thresholds influenced the sorting decisions of the expert system. Experiments with different threshold settings were performed. Lower threshold settings could be used to obtain high quality eggs. This also resulted in more eggs being rewashed, inspected, or rejected. Higher threshold values reduced the number of eggs sorted for rewashing, inspection, or rejection. The threshold variables provided the capability to implement desired sorting policies. The expert system demonstrated significant potential to reduce the work load on human graders.; The color computer vision system, the neural networks, and the expert system formed integral parts of a decision support system for grading eggs. The decision support system was successfully implemented and demonstrated. |
