Evaluation of CNN-Based Models for Classifying Poultry Eggs Images in Variable Angle, Background, and Lighting Conditions

The poultry industry uses expensive scaling machines and human visual methods to sort the weight and defects of eggs. Deep learning models have dived into different types of classification for eggs. Based on state-of-the-art work, models have not yet been trained using a poultry egg image dataset with variability in background, lighting, depth, and angle. Hence, this study aims to evaluate convolutional neural network (CNN) models using a poultry egg image dataset under variable conditions. The main dataset was manually obtained from a poultry farm using mobile phone cameras and contains 6,481 images, which were further organized to classify the size and physical attributes of the eggs. Image augmentation was used for image pre-processing by allowing for random X and Y flips, rotation from -90 to +90 degrees, and scaling from 80% to 120%. Various CNN-based models were used, such as AlexNet, GoogleNet, MobileNetV2, ResNet-50, and SqueezeNet. Results show that the ResNet-based model achieved the highest accuracy of 86.66% for physical attributes. Among the size models, MobileNetV2 scored the highest, with an accuracy of 20.23%. The low accuracy of size models shows that current commonly used CNN models have difficulty classifying images with variable backgrounds, lighting, and camera perspectives.