Evaluation of a Deep Neural Network for Full-Reference Image Quality Assessment in the Context of 3D Data Visualization

Recently, deep Convolutional Neural Networks (CNNs) have proven to provide state-of-theart performance for perceptually accurate Full-Reference Image Quality Assessment (FR IQA). In a natural scene image domain with focus on distortions arising from compression, transmission or image restoration they outperform traditional methods like MSE or PSNR but also metrics like SSIM and FSIM utilizing special hand-crafted features. The reason for this is most likely their explicit training on datasets of this domain. Whether their predictive power is maintained on more shifted image domains is an open question. The
scope of this thesis will be to investigate, how well the performance of CNN-based FR IQA methods is for images showing 2D view projections of 3D renderings together with rendering related image artifacts. For this purpose, the instant3Dhub platform was used to automatically generate images of this target domain and to artificially create rendering artifacts within them. Next, a user study was conducted in order to obtain subjective ground-truth ratings for the distorted images. Performance tests were run exemplarily using the WaDIQaM-FR model. The results show a significant loss of performance on the
target domain. The learning-based approaches, which are strong on the source domain, are perfomance-wise even behind some conventional methods. Thus, it was examined, whether Transfer Learning (TL) and supplying additional information to the network helps enhancing the model performance.