RESUMO
The estimation of image quality and noise perception still remains an important issue in various image processing applications. It has also become a hot topic in the field of photo-realistic computer graphics where noise is inherent in the calculation process. Unlike natural-scene images, however, a reference image is not available for computer-generated images. Thus, classic methods to assess noise quantity and stopping criterion during the rendering process are not usable. This is particularly important in the case of global illumination methods based on stochastic techniques: They provide photo-realistic images which are, however, corrupted by stochastic noise. This noise can be reduced by increasing the number of paths, as proved by Monte Carlo theory, but the problem of finding the right number of paths that are required in order to ensure that human observers cannot perceive any noise is still open. Until now, the features taking part in the human evaluation of image quality and the remaining perceived noise are not precisely known. Synthetic image generation tends to be very expensive and the produced datasets are high-dimensional datasets. In that case, finding a stopping criterion using a learning framework is a challenging task. In this paper, a new method for characterizing computational noise for computer generated images is presented. The noise is represented by the entropy of the singular value decomposition of each block composing an image. These Singular Value Decomposition (SVD)-entropy values are then used as input to a recurrent neural network architecture model in order to extract image noise and in predicting a visual convergence threshold of different parts of any image. Thus a new no-reference image quality assessment is proposed using the relation between SVD-Entropy and perceptual quality, based on a sequence of distorted images. Experiments show that the proposed method, compared with experimental psycho-visual scores, demonstrates a good consistency between these scores and stopping criterion measures that we obtain.
RESUMO
Convolution neural networks usually require large labeled data-sets to construct accurate models. However, in many real-world scenarios, such as global illumination, labeling data are a time-consuming and costly human intelligent task. Semi-supervised learning methods leverage this issue by making use of a small labeled data-set and a larger set of unlabeled data. In this paper, our contributions focus on the development of a robust algorithm that combines active and deep semi-supervised convolution neural network to reduce labeling workload and to accelerate convergence in case of real-time global illumination. While the theoretical concepts of photo-realistic rendering are well understood, the increased need for the delivery of highly dynamic interactive content in vast virtual environments has increased recently. Particularly, the quality measure of computer-generated images is of great importance. The experiments are conducted on global illumination scenes which contain diverse distortions. Compared with human psycho-visual thresholds, the good consistency between these thresholds and the learning models quality measures can been seen. A comparison has also been made with SVM and other state-of-the-art deep learning models. We do transfer learning by running the convolution base of these models over our image set. Then, we use the output features of the convolution base as input to retrain the parameters of the fully connected layer. The obtained results show that our proposed method provides promising efficiency in terms of precision, time complexity, and optimal architecture.
