Investigation on semi-virtual dataset based on semantic segmentation for injection molding process monitoring

Images semantic segmentation on moulds is presented to achieve automatic and continuous defect detection during injection moulding (IM) processing. Generally, a vital prerequisite of training a robust segmentation neural network is to build a suitable dataset, which requires a large amount of image data with plentiful mould colours, different illumination conditions and specific defect labels. The approach requires long data preparation to collect a sufficient dataset, which could meet the needs of online defect detection during injection moulding. To address robust dataset development, the present work uses 3D CAD modelling software to fabricate defects and colours on moulds, creating an augmented dataset mixed with virtual data and real-process data. In this investigation, moulds printed by vat photopolymerization is used as object for which both mould pictures during injection moulding process and virtual mould pictures from 3D modelling software are collected. Mixed datasets with different proportion of virtual images are labelled and fed into the segmentation neural network. The experimental work shows that the proposed method provides a reliable augmentation on the dataset for subsequent IM semantic segmentation framework.