Evaluation of Convolutional Neural Networks for Image Quality Classification based on Synthetic Data

Abstract: In camera production the image quality is of utter importance. Several tests during the production ensure this high quality. In this thesis the possibility of creating a final test, that classifies the image quality with the help of machine learning, specifically convolutional neural networks, was investigated. The data used was made up of synthetic, simulated images with commonly observed quality defects. Eight different network architectures were evaluated on four different types of data sets; two data sets for binary classification, one for multi-class classification, and one data set containing real, non-simulated images. The results were promising with a test accuracy of 0.999 for the binary case with a two-stream network with a DenseNet base. For the multi-class classifi- cation the best test accuracy was 0.989 with the same network. The results showed that there is a high potential for the use of convolutional neural networks for classifying image quality. For a large enough data set a simple convolutional network would be sufficient, achieving similar results as the best network. The networks could handle most of the investigated defects, but seemed to have a problem with blemishes - dust/sensor defects, which is why it would be recommended to have another test for this defect. It was also concluded that there is a need for gathering a larger amount of real images for training since the results on the real data set were at best 0.594 for binary classification. This showed that the networks trained on the simulated images did not translate well to real images.