# A convolutional neural network-based framework for quality control through speckle displacement analysis

## TL;DR

The paper introduces a CNN-based framework to improve quality control by analyzing speckle displacement in images.

## Contribution

A novel CNN architecture is developed to optimize grid overlap sizes for accurate speckle displacement measurement.

## Key findings

- A CNN was optimized using Monte Carlo simulation and grid search for grid overlap calculation.
- The method successfully detected multiple translational movements in simulated and real speckle patterns.
- Validation showed the framework's effectiveness in speckle metrology for quality control.

## Abstract

Among the most advanced techniques for quality control, image processing and optical methods are prominent because of their precision and versatility. These methods often involve analyzing speckles generated by coherent laser illumination because coherent light provides detailed and accurate measurement capabilities. In speckle metrology-based techniques, the accurate measurement of speckle displacements is crucial for detecting faults or deformations in objects. In this study, an advanced algorithm segments the image into overlapping grids, followed by a Fourier-based image registration to accurately quantify the speckle displacements. This method can simultaneously detect multiple translational movements in the different parts of an object. However, proper calculation and assignment of overlap sizes to each grid plays a crucial role in this method, which is where we obtain help from convolutional neural networks (CNNs). We develop a CNN architecture and optimize its hyperparameters using a Monte Carlo simulation algorithm incorporating a grid search and k-fold cross-validation. Finally, we validate the developed method through a case study involving a simulation and real speckle patterns generated by spraying water on a cardboard surface.