Capsule Defect Detection and Classification Using Enhanced MobileNet Model

Abstract

The quality of capsules has a direct impact on human health. However, defects inevitably arise during the capsule manufacturing process, and manual sorting is required. Most studies on capsule defects focus on detecting surface imperfections and those on internal capsule defects remain relatively limited. This study classifies four types of capsules using the deep learning model MobileNet, aiming to accurately identify surface and internal defects. A total of 2872 capsule images are used to evaluate the model's classification performance prior to and following optimization. The MobileNet model categorizes capsule images into four types: Normal capsules, deformed capsules, impure capsules, and bubbled capsules. Stratified cross-validation is applied to partition the dataset into 80% training, 10% validation, and 10% testing sets for ten-fold cross-validation. The model's performance is evaluated using three metrics: Precision, recall, and F1-score. The results are compared with two classic deep learning models, four traditional machine learning models (VGG16, ResNet101, KNN, and SVM), as well as decision trees and random forests. The findings demonstrate that the MobileNet model exceeds the performance of the other models, achieving precision, recall, and F1 scores of 94.24, 94.75 and 94.23%, respectively. Through transfer learning and improving the top layers of the MobileNet model with dropout, L2 regularization, Batch Normalization (BN), and average pooling, the model accuracy is improved by 7.95%, indicating promising performance and potential in detecting capsule defects.