Mask R-CNN for real-time parts classification on an enamel paint coating line

Abstract

An enamel paint coating line has relied heavily on human intervention and observation to determine the total number of parts produced in one shift. This was done by manually counting the parts as they were taken off the conveyor line which has led to inaccurate or lately received production performance data. With a manufacturing process that has a changing number of classes and a variation of the part’s physical orientation within the process, the importance of investigating an object detection model was realized. The rising demand to attain live production data has added more importance to monitoring and reporting within the industrial automation sector. Real-time parts screening requiring human intervention for data input may not be a feasible solution within a fast-moving consumer goods manufacturing facility. This has led to the proposed study using Mask Region-based Convolutional Neural Network (R-CNN) to detect objects within an image. The state of research surrounding object detection has increased rapidly over the past decade with several models being developed and adapted for industrial automation applications. However, there has been limited research regarding models that can perform instance segmentation to provide production performance data. This has prompted the study within this field. The objective of the study is to classify and accurately report on manufactured parts identified on an enamel paint-coating conveyor line. At any given instant, the parts may not be in the exact coordinates within the desired area of interest and the classes of objects may vary based on changing production requirements. To mitigate these challenges, this study proposes the use of a trained Mask R-CNN model to detect the objects and their associated class. Images are acquired using a high-definition camera fixed to a position next to the enamel coating line. This study compared the average precision based on changes made to the learning rate and the intersection over union (IoU) thresholds of the model. The outcome was analysed using the precision-recall curve and the confusion matrix to determine the highest average precision and overall accuracy. The highest achieved average precision obtained from the model was 98.27% with an overall accuracy of 98.24% using real-time captured images of the manufactured parts. The results satisfied the acceptable standard for the average precision of 97.5% set by the plant production quality engineers. Future research would include the use of pixel-wise segmentation generated from the mask branch in determining the objects’ exact location, this would aid in efficient robotic spray positioning.