Duck Egg Quality Classification Based on its Shell Visual Property Through Transfer Learning Using ResNet-50.

Duck egg quality inspections are done manually in the Philippines which is not reliable due to human subjectivity, visual stress, and tiredness. There is no documentation regarding the general standard on determining the quality of duck eggs and local farmers have different standards. This research aims to classify duck egg quality into 3 classes namely, Balut/Penoy, Salted Egg, and Table egg. Two angles of 600 duck eggs were captured inside an image acquisition setup. Using ResNet-50 as a base model, its last fully connected layer was replaced by a classifier block. Hyperparameter tuning with Stratified 5-fold Cross Validation was utilized. It was observed that batch size of 8, epoch of 110, and learning rate of 0.0001 has given the lowest validation loss which was used to train the final model. Performance Metrics was obtained. Overall, the result of the model yielded 90%, 95%, and 65% for Balut/Penoy, Salted Egg, and Table egg, respectively, averaging an 83.33% overall accuracy of the model. As observed, the model is not able to accurately differentiate between hairline and broken cracks. Additionally, falsely classified images also occur when the size of an egg is close to the threshold to other sizes. [ABSTRACT FROM AUTHOR]