Poultry Egg Classification System Using Deep Learning

Abstract

This research proposes the use of CNN in the classification systems of unwashed eggs. In this work, a database of 1500 images is generated from unwashed brown eggs, which includes intact, cracked, dirty, and dirty-cracked classes, to make a comparison between the ResNet34, ResNet50, and VGG19 models using two batch sizes, 32 and 64, during training. Additionally, a custom learning rate assignment was applied for the pre-trained layers and classification layers, with rates of 0.001 and 0.0001 for the respective layers. The models were trained with retrained weights and followed a fine-tuning configuration of 20% of the layers. The highest accuracy of 97.33% was achieved with the VGG19 model using a batch size of 64.