Dataset: Natural and synthetic datasets for rapid deep-learning-based optical measurement of printed linear structures

Conventional optical measurement techniques are beneficial in manufacturing processes due to their fast and non-intrusive operation. However, they require sophisticated and expensive equipment as well as increased personnel qualification. While the integration of machine learning contributes to alleviating these requirements, it needs a time-consuming and tedious preparation of training datasets. We introduce a twin concept to conduct rapid measurement of various geometrical properties of printed lines using deep learning. For the first time, a tedious collection and assessment of training data is substituted by synthetic generation of digital samples, which enables flexible, fast, and precise data processing. A full-field analysis of a non-preprocessed image conducted by a neural network facilitates measurement of multiple geometrical properties of an object at once. Corresponding network architecture, workflows, and metrics are outlined. The application area of the concept is demonstrated, but not limited to the field of functional printing. The method can easily be tailored to a wide range of engineering fields.

Conventional optical measurement techniques are beneficial in manufacturing processes due to their fast and non-intrusive operation. However, they require sophisticated and expensive equipment as well as increased personnel qualification. While the integration of machine learning contributes to alleviating these requirements, it needs a time-consuming and tedious preparation of training datasets. We introduce a twin concept to conduct rapid measurement of various geometrical properties of printed lines using deep learning. For the first time, a tedious collection and assessment of training data is substituted by synthetic generation of digital samples, which enables flexible, fast, and precise data processing. A full-field analysis of a non-preprocessed image conducted by a neural network facilitates measurement of multiple geometrical properties of an object at once. Corresponding network architecture, workflows, and metrics are outlined. The application area of the concept is demonstrated, but not limited to the field of functional printing. The method can easily be tailored to a wide range of engineering fields.

The archive includes two datasets. The dataset of 20,000 synthetic images was generated for the neural-network training. Filename structure: [sequence number]-[edge class]-[line width in mpx]-[SD in mpx]-[contrast level]-[noise level]. The dataset of 200 natural images was collected to test pre-trained model. Filename structure: [line type, code]-[line width in nm]-[SD in nm]-[pixel density in mpx/mm]. Pixel density is reciprocal to the image resolution. Line type and code structure: [diw/ijp]: direct-written or inkjet-printed line, [first number]: sequence number, [a/b]: printed line or interline space, [last number]: 0 - original image, 1 - vertically mirrored and gamma-corrected image, 2 - horizontally mirrored and gamma-corrected image, 3 - 180° rotated and gamma-corrected image, 4 - solely gamma-corrected image. Gamma correction was randomly set in the range [0.5, 1.5].