Abstract:
Abstract: Moldy core in apple (a common internal disease infected by fungal) has widely resulted in quality loss and food safety for fruits and by-products, such as concentrated juice and cider. However, the slightly infected apples are difficult to be picked out, because there are no visible symptoms in the fruit appearance at present. Traditional manual inspection is highly destructive, subjective, and time-consuming, due mainly to require cutting apples into halves for the visual evaluation in the presence or absence of internal defects. In addition, little research has been focused on the early detection of internal disease in fruits. Consequently, there is an urgent demand to develop the nondestructive detection system for the early detection of moldy apple core. Therefore, it is an urgent demand to develop a nondestructive early detection for the apples with a moldy core. In this study, a nondestructive vibro-acoustic setup was employed to detect the apples with slight moldy-core using two identical piezoelectric transducers. The obtained vibro-acoustic signals were transformed to the images using Symmetrized Dot Pattern (SDP) algorithm. SDP images were then used to extract the depth feature using the transfer learning of three Convolutional Neural Networks (CNNs), including AlexNet, VGG16, and ResNet50. The extracted features were fed to train the Support Vector Machine (SVM) classifier, finally to identify the slightly moldy apple core (moldy-core degree less than 7%). Specifically, the largest difference of shape feature was found among the SDP images of sound and moldy-core apples, when the time lag coefficient l was 25 and the angular gain factor was 50o. As such, various SVM classification models were constructed in this case using the different CNN structures and kernel functions. Correspondingly, the ResNet50-SVM-gaus model performed the higher classification in the training set with less training time and the number of parameters, compared with the AlexNet-SVM-line model. Subsequently, the super parameters were selected to optimize the network structure in the trained ResNet50-SVM-gaus model, including learning rate and epochs. Specifically, the classification accuracy of model was improved from 91.38% to 99.63%. Furthermore, the total classification accuracy of the model in the test set reached 96.97% using an imbalanced dataset with the sound to diseased apples of 10:1. Meanwhile, the Stable Precision (SP), Stable Recall (SR), Stable F1-score (SF), Kappa Coefficient (KC), and Matthews Correlation Coefficient (MCC) of the ResNet50-SVM-gaus model were 80.19%, 90.36%, 86.21%, 82.54%, and 82.68%, respectively. These indicated that the ResNet50-SVM-gaus model achieved the accurate classification for the early detection of apples with a slight moldy core. Therefore, the ResNet50-SVM-gaus model can be expected to enhance the classification performance for the minority variety of moldy-core apples in the early stage. Consequently, the vibro-acoustic approach combined with SDP demonstrated a promising potential to early detect fungal diseases in moldy apple core. The finding can provide the theoretical reference for the early detection of diseases inside the fruits.