Abstract:
Changing in lifestyle, the perceived impurity of tap water, and the perceived cleanliness of
bottled water can be considered as the cause for the increasing demand of plastic water
bottle users. Bottling companies strive for quality when producing bottled water. Because,
different defects occurred on the bottles in the production process. So, it’s important to
check defects during the water bottling process. It is a difficult task for human beings to
detect the defects from a huge number of water bottles in the production line during the
inspection process. Therefore, various researchers are motivated to develop defect
detection models based on image processing. However, as we have investigated from the
previous research they are motivated to detect different defects between glass and plastic
bottles. But the previous researchers do not consider different shape of the bottles. In this
study, we have performed a plastic water bottle defect detection using a machine learning
approach. We have examined defects such as (broken cap defect, no cape defect, scratch
cap defect, shape defect, and water level defect) and non-defects of the bottle. To achieve
the objective of the study, plastic water bottle images with different shapes and sizes are
considered from Choise, Eleleta, and Kefita plastic water bottling PLCs. A total number of
9,000 plastic water bottle images were taken by using Samsung Galaxy S7. After getting
the dataset we have done preprocessing tasks like resizing, noise reduction, grayscale
conversion, histogram equalization, and segmentation. For feature extraction, we have used
Histogram Oriented Gradients (HOG), Convolutional Neural Network (CNN), and a
combination of HOG and CNN. For classification, we have used end-to-end CNN, Support
Vector Machine (SVM), and K-Nearest Neighbor (KNN). The proposed hybrid CNN and
HOG feature trained with CNN, SVM, and KNN classifiers achieved the highest accuracy
obtained in this work. From the experimental results, our end-to-end CNN model achieved
94.33% accuracy, using SVM classifier with HOG, CNN, and combined (i.e. HOG and
CNN) features is 82.27%, 94%, and 99% respectively. And also we have got an accuracy
of using KNN classifier with HOG, CNN, and combined (i.e. HOG and CNN) features is
81%, 82%, and 97% respectively. Therefore, the combined feature of HOG and CNN with
SVM classifier has outperformed the better accuracy for detecting defects in plastic water
bottles.
Keywords: Histogram Oriented Gradient, Convolutional Neural Network, Defect
Detection, Feature extraction, Support Vector Machine, and K-Nearest Neighbor.