摘要：AbstractThis paper aims to help improve the creation of deep learning-based vision systems for consumer goods manufacturing at Procter and Gamble (P&G) using smart, realistic synthetic data creation. This synthetic data creation is based on a novel data resampling technique that utilizes ordinal class information to create hard-to-capture minority class defects, which has been appropriately name Class Ordered Minority Oversampling Technique (COSMOTE), while also minimizing the overall data collection efforts, which can be a costly and disruptive process to the plants themselves. A particular challenge to these applications is the small number of samples that may be captured for defective products, especially when changes in the process or artwork are made. By leveraging these ordinal quality classes, the information from the classes themselves can enable a minimal training dataset size for faster start to finish model development. A brief literature review of existing resampling techniques is provided to highlight the gaps in these sparse sample use cases and the workflow to generate and validate these synthetic images is also outlined. This paper explains the benefits of intelligent synthetic data creation within this particular manufacturing space by addressing both data imbalance and sparse sample datasets.
