Deep Learning-Based Visual Inspection of Straw Defects in Beverage Box Packaging for Process Quality Monitoring

Visual inspection is an important part of packaging quality control, especially when small external components, such as attached straws, can affect product presentation and consumer acceptance. In beverage box packaging, defects such as bent, loose, missing, or pierced straws may be overlooked during repetitive manual inspection. To support process quality monitoring, this study presents a deep learning-based visual inspection approach for classifying straw defects using the publicly available Juice Box Straw Defects Dataset. The dataset contained 2000 images equally distributed across five classes: Bent, Good, Loose, Missing, and Pierced. Three Convolutional Neural Network (CNN) architectures, namely DenseNet121, Xception, and InceptionV3, were evaluated using a transfer learning strategy. Each model was first trained with a frozen pretrained backbone and then fine-tuned by unfreezing selected final layers to improve task-specific feature learning. Model Performance was assessed using accuracy, precision, recall, F1-score, and confusion matrix analysis. Baseline results showed moderate classification performance, with InceptionV3 obtaining the highest baseline accuracy of 0.7433. After fine-tuning, all models improved substantially, with fine-tuned Xception achieving the best test performance, recording 0.9633 of accuracy, 0.9637 of precision, 0.9633 of recall, and 0.9632 of F1-score. Grad-CAM visualization further indicated that the model focused on straw-related regions during classification. These findings suggest that fine-tuned CNN models can provide effective image-based support for straw defect inspection in beverage box packaging.