Evaluation of Different Deep Learning Network Architectures in Egg Freshness Detection Based On Sound Signals

Food security, which is directly related to the health of people, has always been a concern of all nations. Eggs are consumed in many food industries and are in the daily diet of many people, so detection their freshness is very important. In this study, the capability of the acoustic system as a non-destructive method for egg freshness detection was investigated. Samples were stored at room temperature for 1, 4, 7, 10, 13 and 16 days. After data collection, all audio signals were converted to images, using spectrogram. In this study, the freshness of samples was evaluated using two criteria; Haugh unit and air cell height as a destructive test. The results of destructive test showed that all samples stored for 16 and 13 days and also 80% of samples stored for 10 days faced with quality losses during storage. According to grading criteria, these samples were considered as unfresh eggs. Therefore, the samples were divided into two groups: fresh eggs (stored for 1, 4 and 7 days) and unfresh (stored for 10, 13 and 16 days). Four pre-trained deep learning networks AlexNet, VGGNet, GoogLeNet and ResNet were used in this study among which ResNet had the best classification accuracy with an average of 71.5%.