Homepage

Description

This work aims at analyzing the statistical properties of computer generated images by application of the Natural Scene Statistics (NSS) models. For the purpose of validation of the usefulness of NSS in this domain, we have compiled the ESPL synthetic image database, which contains high quality synthetic color images from the Internet, mostly 1920 x 1080 pixels.The images are primarily chosen from popular video games and animation movies. Some video games which are considered are multiplayer role playing games (for example, War of Warcraft, MegaGlest), first person shooter games (such as Counter Strike), bike and car racing games, and games with more realistic content, such as FIFA. Some of the animation movies, from which the images were collected are, The Lion King, the Tinkerbell series, Avatar, The Beauty and the Beast, Monster series, Ratatouille, the Cars series etc.Care has been taken to provide as varied a range of interest as possible, by incorporating both natural and non-photorealistic renderings of human figures, human-made objects, fantasy figures such as fairies and monsters, close up shots, wide angle shots, images showing both high and low degrees of color saturation, and background textures with no foreground object.In addition to the pristine images, we have considered some distortions, in order to simulate the artifacts introduced to the image while rendering, or transmission over a network. Each image has been degraded using five levels of each type of distortion, ranging from just visible artifacts, to the artifacts which are visible to a large extent.Some of the distortions considered are described below:High Frequency noise:Noise is perhaps the most pervasive distortion present in images, be it natural or synthetic. The synthetic image may get corrupted with noise, either during rendering (for example, in case of any random sample based methods, such as photon mapping based on Monte-Carlo techniques), or during transmission over a network. Three types of noise, namely Gaussian noise, Salt-and-Pepper noise, and Speckle noise, have been considered.Interpolation:Artifact results from insufficient super sampling, or lack of an anti-aliasing filter, and is caused by downsampling the reference image, and upsampling to the original resolution using an interpolation method. These effects are primarily visible at edges, such as shadow maps, and the users see them as jaggedness.Banding:Banding artifact is an unintended side effect of color quantization, which can affect both natural and synthetic images. Sometimes, the bit depth might be insufficient to accurately sample a continuous gradation of a color tone. Hence, this continuous gradation might appear as a series of discrete bands of color. This becomes especially prominent in the large smooth textureless regions of the rendered image, such as the sky.Ringing:Ringing artifact is a type of low frequency noise, which arises as a spurious signal or "ghosting" at the edges in the images. For this database, this artifact was introduced by filtering the image with a two-dimensional sinc function, with varying frequencies, and truncated to a certain window size. Low frequency noise can arise in many photon mapping processes.Gaussian Blur:Images may be blurred as an effect introduced by many denoising techniques, which tend to remove the Gaussian noise with low pass filtering with a Gaussian kernel. In synthetic images, in order to give the shadows a realistic soft edge, they might be filtered using a Gaussian kernel, but overdoing the filter operation can introduce visible blurring distortions.JPEG compression:The blocking artifacts appear when a low compression ratio is chosen for JPEG coded images, especially under low network bandwidth conditions. This artifact can be encountered in the scenario of cloud gaming, where the rendered game image might be transmitted to the users playing using a low bandwidth network.Related publications:D. KunduandB. L. Evans,"Spatial Domain Synthetic Scene Statistics",Proc., Asilomar Conf. on Signals, Systems, and Computers, Nov. 2-5, 2014, Pacific Grove, CA USA.D. KunduandB. L. Evans,"Full-Reference Visual Quality Assessment for Synthetic Images: A Subjective Study",Proc. IEEE Conf. on Image Processing, Sep. 27-30, 2015, Quebec City, Canada, submitted.D. KunduandB. L. Evans,""No-reference Synthetic Image Quality Assessment using Scene Statistics",Proc. IEEE Conf. on Image Processing, Sep. 27-30, 2015, Quebec City, Canada, submitted.

Related Papers

• JPEG compression: The blocking artifacts appear when a low compression ratio is chosen for JPEG coded images, especially under low network bandwidth conditions. This artifact can be encountered in the scenario of cloud gaming, where the rendered game image might be transmitted to the users playing using a low bandwidth network. [link]
• Gaussian Blur: Images may be blurred as an effect introduced by many denoising techniques, which tend to remove the Gaussian noise with low pass filtering with a Gaussian kernel. In synthetic images, in order to give the shadows a realistic soft edge, they might be filtered using a Gaussian kernel, but overdoing the filter operation can introduce visible blurring distortions. [link]
• Ringing: Ringing artifact is a type of low frequency noise, which arises as a spurious signal or "ghosting" at the edges in the images. For this database, this artifact was introduced by filtering the image with a two-dimensional sinc function, with varying frequencies, and truncated to a certain window size. Low frequency noise can arise in many photon mapping processes. [link]
• Banding: Banding artifact is an unintended side effect of color quantization, which can affect both natural and synthetic images. Sometimes, the bit depth might be insufficient to accurately sample a continuous gradation of a color tone. Hence, this continuous gradation might appear as a series of discrete bands of color. This becomes especially prominent in the large smooth textureless regions of the rendered image, such as the sky. [link]
• Interpolation: Artifact results from insufficient super sampling, or lack of an anti-aliasing filter, and is caused by downsampling the reference image, and upsampling to the original resolution using an interpolation method. These effects are primarily visible at edges, such as shadow maps, and the users see them as jaggedness. [link]
• High Frequency noise: Noise is perhaps the most pervasive distortion present in images, be it natural or synthetic. The synthetic image may get corrupted with noise, either during rendering (for example, in case of any random sample based methods, such as photon mapping based on Monte-Carlo techniques), or during transmission over a network. Three types of noise, namely Gaussian noise, Salt-and-Pepper noise, and Speckle noise, have been considered. [link]