Computer vision-based method of pre-alignment of a channel optical waveguide and a lensed fiber

Abstract

The work is devoted to a technique of pre-alignment of a lensed fiber and a channel waveguide of a photonic integrated circuit using computer vision methods. The design and main units of a machine vision system with illumination of the adjusted objects in reflected light are described. The technique includes detection of the spatial position of the end face of the photonic integrated circuit, fixed at an angle of 90 ± 1° to the horizontal axis of the frame, detection of the coordinates of the end face of the lensed fiber, and the subsequent correction of the position of the lensed fiber using a manipulator system. We propose a method of searching and selecting a single line corresponding to the end face of a photonic integrated circuit using a Hough transform; methods for grouping discontinuous contours of the lensed fiber and true contour determination. These methods are based on a priori knowledge of the lens geometry. Also, we describe options for suppressing noise and overcoming various defects in images. It has been shown experimentally that the error of angle determination of a lensed fiber depends on the distance between the lens and the end face of the photonic integrated circuit. The presented technique makes it possible to determine the longitudinal and angular displacements between the fiber lens and the end face of the photonic integrated circuit with errors less than 4 μm and 0.05°, respectively.

Digital Twins for Lighting Analysis: Literature Review, Challenges, and Research Opportunities

“…Even though in real-world applications this approach is acceptable, because a filter with specified parameters is often used, it is necessary to be able to customize the filter parameters when creating and optimizing prototypes of the final system. For example, Karnaushkin and Sklyarenko (2022) used Canny edge detector to develop a computer vision-based method of pre-alignment of a channel optical waveguide and a lensed fiber, and the values of the algorithm parameters were being tuned during the development stage. Sometimes it is necessary to give a user the ability to specify the system parameters, as in the case of an ANN-based image pre-compensation system, which must adjust to the parameters of a particular person's eye (Yu et al 2021) and thus cannot be retrained for each possible set of parameters.…”

“…But there is still a lack of studies about the neural networks capability of approximating algorithms, particularly image processing algorithms. Many convolutional network architecture families, such as ConvNet, ResNet, UNet, etc., are successfully used in various image processing tasks (Li et al 2021;Andreeva et al 2019), including ones with existing algorithmic solutions (Karnaushkin and Sklyarenko 2022;Panfilova and Kunina 2020). However, it is not obvious that especially for tasks where specialized algorithmic solutions are known, a general-purpose neural network could solve the same task with comparable computational and representational efficiency.…”

ECMS 2022 Proceedings edited by Ibrahim A. Hameed, Agus Hasan, Saleh Abdel-Afou Alaliyat

“…Even though in real-world applications this approach is acceptable, because a filter with specified parameters is often used, it is necessary to be able to customize the filter parameters when creating and optimizing prototypes of the final system. For example, Karnaushkin and Sklyarenko (2022) used Canny edge detector to develop a computer vision-based method of pre-alignment of a channel optical waveguide and a lensed fiber, and the values of the algorithm parameters were being tuned during the development stage. Sometimes it is necessary to give a user the ability to specify the system parameters, as in the case of an ANN-based image pre-compensation system, which must adjust to the parameters of a particular person's eye (Yu et al 2021) and thus cannot be retrained for each possible set of parameters.…”

“…But there is still a lack of studies about the neural networks capability of approximating algorithms, particularly image processing algorithms. Many convolutional network architecture families, such as ConvNet, ResNet, UNet, etc., are successfully used in various image processing tasks (Li et al 2021;Andreeva et al 2019), including ones with existing algorithmic solutions (Karnaushkin and Sklyarenko 2022;Panfilova and Kunina 2020). However, it is not obvious that especially for tasks where specialized algorithmic solutions are known, a general-purpose neural network could solve the same task with comparable computational and representational efficiency.…”