Machine learning approaches for quality control in pulp packaging manufacturers

In molded pulp packaging manufacturing, defect detection and classification processes are critical to ensuring the products meet quality criteria. Yet most manufacturers still rely on human-based manual visual defect classification which can be inconsistent and labor intensive. In this research, we introduce the conjunction of machine vision hardware and machine learning to build a conceptual framework for an automated molded pulp packaging defect detection system. The conceptual framework consists of two modules. First, the image acquisition module setups appropriate hardware and configuration such that high-quality images can be acquired. The second is a machine learning module that constructs a deep learning model with hyper-parameter tuning to automatically detect the defects on the surface of molded pulp products. Our proposed model is based on deep learning model - the Xception architecture, which is recently developed and expected to be more robust on defect detection. In comparison with Traditional machine learning algorithms - SVM and Naive bayes have been widely used in the field of industrial detection. The oriented FAST and rotated BRIEF (ORB) and Bag-of-Visual-Word (BoVW) are implemented for pre-feature extraction. Since molded pulp packaging has obstacles on surface fluctuation by color, grain pulp fiber and non-repeating defect pattern, the Negative Monochrome (NGMC) image preprocessing is proposed to enhance the visibility of defects on the surface and reduce undesired features. The extracted features must be able to describe and distinguish images categories, which could be a limitation for traditional algorithms that required pre-feature extraction. The results demonstrate that the Xception model trained with NGMC images resolution 192x192 and learning rate 0.001 achieved more than 92.98% accuracy and best generalize across datasets from different production lots, which suggests that the robustness of our conceptual framework has the potential to be utilized in industrial applications.