Detecting Industrial Potato Chips Defects with Machine Learning Methods Using Deep Features

Abstract

Detection of production defects in industrial foods is vital to protect consumer health. Early detection of these errors can minimize the economic losses of manufacturers by reducing the costs of recalls and production stoppages. Additionally, continuous monitoring and improvement of product quality can increase brand reliability and customer satisfaction. Image processing can detect product defects, minimize human error and increase efficiency by performing uninterrupted inspection on the production line. Based on these reasons, this study aimed to detect potato chip errors with image processing. PepsiCo Lab Potato Chips Quality Control image dataset was used in the study. There are two classes in the dataset: defective and not defective. There are 967 images in total. SqueezeNet Convolutional Neural Network (CNN) architecture was used to extract the features of the images. With this architecture, 1000 features obtained for each image were classified with Artificial Neural Network (ANN), K Nearest Neighbor (KNN), Random Forest (RF) machine learning methods. As a result of the classifications, 0.986 classification accuracy was obtained from the ANN model, 0.927 from the KNN model, and 0.962 from the RF model. F1 Score, precision, recall and specificity metrics were used to compare the models in detail. According to the data obtained from the experimental results, it is predicted that the proposed feature extraction and classification models can detect industrial production errors occurring in potato chips.Detection of production defects in industrial foods is vital to protect consumer health. Early detection of these errors can minimize the economic losses of manufacturers by reducing the costs of recalls and production stoppages. Additionally, continuous monitoring and improvement of product quality can increase brand reliability and customer satisfaction. Image processing can detect product defects, minimize human error and increase efficiency by performing uninterrupted inspection on the production line. Based on these reasons, this study aimed to detect potato chip errors with image processing. PepsiCo Lab Potato Chips Quality Control image dataset was used in the study. There are two classes in the dataset: defective and not defective. There are 967 images in total. SqueezeNet Convolutional Neural Network (CNN) architecture was used to extract the features of the images. With this architecture, 1000 features obtained for each image were classified with Artificial Neural Network (ANN), K Nearest Neighbor (KNN), Random Forest (RF) machine learning methods. As a result of the classifications, 0.986 classification accuracy was obtained from the ANN model, 0.927 from the KNN model, and 0.962 from the RF model. F1 Score, precision, recall and specificity metrics were used to compare the models in detail. According to the data obtained from the experimental results, it is predicted that the proposed feature extraction and classification models can detect industrial production errors occurring in potato chips.