Image Processing and Neural Network Optimization Methods for Automatic Visual Inspection

Abstract

In intelligent production lines, methods of automatic visual inspection are used to continuously record process parameters and production results. Using the quality control of thin film solar modules as an example, this paper shows how visual inspections can be automated by using neural networks. The starting point of this automation is an image of a manufactured solar module generated by the inverse operation of the solar cell and the associated electroluminescence.

It turns out that the amount of data generated with every image is very high despite truncation of irrelevant areas and a reduction of the resolution from 1024 x 1024 pixels to 256 x 128 pixels. Without further preprocessing of the data, A neural network would be built that would have 32768 input nodes due to the pixels acquired. The Convolutional Neural Networks (CNNs) are usually considered for such image classification problems. However, their use increases the complexity of the architecture of the network and thus the number of parameters to be optimized. Therefore, in addition to automated visual inspection, this work addresses the question of how image processing methods can be used for high-quality and efficient implementation.

The Fast Fourier Transform is used for data preprocessing to enable the use of a multilayer perceptron (MLP) rather than a CNN. It is shown that the computation time can be reduced by a factor of 13 by image preprocessing and using the Fast Fourier Transform to compress the dataset. The reduced computation time is a prerequisite for optimizing the neural network hyperparameters. Particle-Swarm Optimization and Genetic Algorithms are implemented and compared to perform a Neural Architecture Search (NAS) for a MLP and to optimize the other hyperparameters. The methods lead to architectures with which an accuracy of more than 99 % is achieved. The high accuracy of the presented method recommends it for further projects of automatic visual inspection. The method thus allows digitalization in statistical process control and thus contributes to the implementation of the Quality 4.0 vision.