Non-destructive rapid prediction of raw salted duck egg quality using optimized convolutional neural network

Abstract

A system to determine the quality of salted duck eggs is necessary to increase egg production and prevent consumption of bad-salted duck eggs. Egg images were captured by a camera with an LED light, and a box with a black interior was utilized. A proposed optimized CNN (opti-CNN) model using a small number of parameters (6.5 million) and convolutional layers was developed to recognize the key elements of the salted duck egg image utilizing pooling, activation, fully connected layers, and a sigmoid layer to enhance the performance of the CNN model. VGG-16, AlexNet, Inception-V3, and MobileNet were used for comparison. The findings revealed that all the models had accuracy levels above 90%. The MobileNet model exhibited 100% accuracy, which was the highest, and the opti-CNN exhibited 97% accuracy. However, the opti-CNN showed a computation time of only 50 ms, whereas the other models showed a computation time above 500 ms. These findings show that determining the quality of salted duck eggs using the proposed opti-CNN yields outstanding results and can be applied in large-scale sorting procedures.