Defect detection in the food industry through deep learning models

In this thesis study, the effectiveness of deep learning-based image processing methods for automatic detection of defective and non-defective samples based on visual characteristics prior to entering the production line in the packaged food industry was investigated. Within the scope of the study, 405 defective and 369 non-defective product images obtained from the production line were preprocessed and standardized only in terms of format to make them suitable for modeling. During the modeling process, a total of 10 different deep learning architectures Inception, InceptionResNetV2, DenseNet121, EfficientNetB0, EfficientNetV2B0, MobileNet, ResNet50, VGG16, VGG19, and Xception were applied. All models employed a 5-fold cross-validation method, and performance evaluations were conducted accordingly. The success of each model was analyzed in detail by comparing training and test losses as well as metrics including accuracy, precision, recall, and the F1 score