Optical Inspection Methods for the Quality Verification of Printed Circuit Boards and Polymer Components

Quality inspection is an important part of industrial production. It is used to determine the quality of manufactured products and is usually carried out by optical inspection. Due to increasing competition between companies, the requirements for quality inspection are becoming ever more stringent, which has led to more attention being paid to scientific research in this area. The application of deep learning methods in optical inspection has further increased interest. Optical inspection with deep learning methods is divided into a hardware part and a software part. In this thesis both parts are investigated. A high-resolution inspection concept for the inspection of printed circuit boards (PCBs) arrays was developed, which detects defects in the micrometer range on the gold areas on PCBs, which are highly reflective. Subsequently, Photometric Stereo methods for the visualization of topographical defects on PCBs were investigated. Topographical defects are unwanted anomalies that appear either as a dent or a peak on the surface.
Non-topographical defects, on the other hand, are characterized by a different color compared to the rest of the surface. Traditional Photometric Stereo and an existing Neural Inverse Rendering framework were used to visualize the topographic defects. The results showed that the defects are very well recognizable in the normal map. In the method with the Neural Inverse Rendering framework, the defects could also be displayed in a depth map. However, the results with this method were only partially reproducible, as the used framework was not sufficiently adapted to the inspection environment. The detection and classification of topographical and non-topographical PCB defects was performed using a Convolutional Neural Network (CNN). The input to the CNN was a 6-channel image consisting of the normal map and a mean image. The mean image is calculated from the average of the Photometric Stereo images and visualizes the nontopographical PCB defects. With the trained CNN, a prediction accuracy of 95 % could be achieved on the test dataset. When training CNNs, hyperparameters are defined at the beginning, which form a component in achieving optimal prediction accuracy. In this work, the prediction accuracy on an industrial dataset was investigated by varying the hyperparameters batch size and learning rate. The results show that with small batch sizes and small learning rates the highest F1-score and True Positive Rate could be achieved. The performance of training CNNs heavily depends on the dataset. The model learns to predict the component quality based on the training dataset. If there are deviations between the training and test datasets, the prediction accuracy is reduced. The generalization capability of CNNs describes the performance of CNNs on unseen test data. In this work, the generalization ability of a pre-trained and a non-pre trained CNN in predicting topography parameters under deviations between training and test dataset was investigated. The dataset used consists of component series produced using foam injection molding. By varying the mold temperature, changes in the surface topography could be achieved. The components have six different surface structures. Topography parameters of each of these surface structures were measured for the dataset using an optical microscope. A special training strategy was used to test the generalization capability. As a result, the predictions were most accurate for both the pre-trained and non-pre-trained CNNs at higher mold temperatures. At lower mold temperatures and a compact molded part series, the determination of the topography parameters was the least accurate.