Few-shot incremental learning in the context of solar cell quality inspection

Deep Neural Networks have shown high defect detection rates in industrial setups, surpassing other more traditional manual feature engineering-based proposals. This has been mainly achieved through supervised training, where a great number of annotated samples are required to learn good classification models. However, obtaining such a large amount of data is sometimes challenging in industrial scenarios, as defective samples do not occur regularly, and certain types of defects only appear occasionally. In this work, we explore the technique of weight imprinting in the context of solar cell quality inspection. This technique allows to incorporate new classes into the classification network using just a few samples. We tested the technique by first training a base network for the segmentation of three base defect classes and then sequentially incorporating two additional defect classes. This resulted in a network capable of segmenting five different defect classes. We also experimented with the network architecture, resulting in more precise segmentation and defect detection results. The experiments’ results have shown that this technique allows the network to extend its capabilities regarding the detection of new defect classes using just a few samples, which can be interesting for industrial practitioners.