Image-based characterization of laser scribing quality using transfer learning

Ultrafast laser scribing provides a new microscale materials processing capability. Due to the processing speed and high-quality requirement in modern industrial applications, it is important to measure and monitor quality characteristics in real time during a scribing process. Although deep learning models have been successfully applied for quality monitoring of laser welding and laser based additive manufacturing, these models require a large sample for training and a time-consuming data labelling procedure for a new application such as the laser scribing process. This paper presents a study on image-based characterization of laser scribing quality using a deep transfer learning model for several quality characteristics such as debris, scribe width, and straightness of a scribe line. Images taken from the laser scribes on intrinsic Si wafers are examined. These images are labelled in a large and a small dataset, respectively. The large dataset includes 154 and small dataset includes 21 images. A novel transfer deep convolutional neural network (TDCNN) model is proposed to learn and assess scribe quality using the small dataset. The proposed TDCNN is able to overcome the data challenge by leveraging a convolutional neural network (CNN) model already trained for basic geometric features. Appropriate image processing techniques are provided to measure scribe width and line straightness as well as total scribe and debris area using classified images with 96 percent accuracy. Validating model's performance based on the small data set, the model trained with the large dataset has a similar accuracy of 97 percent. The trained TDCNN model was also applied to a different scribing application. With 10 additional images to retrain the model, the model accuracy performs as well as the original model at 96 percent. Based on the proposed TDCNN classification of debris on a scribed image of straight lines, two algorithms are proposed to compute scribe width and straightness. The results show that all the three quality characteristics of debris, scribe width, and scribe straightness can be effectively measured based on a much smaller set of images than regular CNN models would require.

Automated summary

This paper presents a study on image-based characterization of laser scribing quality using a deep transfer learning model. By training on a small dataset, the proposed TDCNN model is able to effectively measure debris, scribe width, and scribe straightness, achieving an accuracy of 97% on the large dataset.