Steering Image Generation with Wavelet Based Perceptual Metric

Abstract

It is often the case that images generated by image synthesis algorithms are judged by visual examination. The user resorts to an iterative refinement process of inspection and rendering until a satisfactory image is obtained. In this paper we propose quantitative metrics to compare images that arise from an image synthesis algorithm. The intent is to be able to guide the refinement process inherent in image synthesis. The Mean-Square-Error (MSE) has been traditionally employed to guide this process. However, it is not a viable metric for image synthesis control. We propose the use of a wavelet based perceptual metric which incorporates the frequency response of the Human Visual System. A useful aspect of the wavelet based metric is its ability to selectively measure the changes to structures of different sizes and scales in specific locations. Also, by resorting to the use of wavelets of various degrees of regularity, one can seek different levels of smoothness in an image. It is rare that such level of control can be obtained from a metric other than a wavelet based metric. We show the usefulness of our metric by examining it's effectiveness in providing insights for common operations of an image synthesis algorithm (e.g., blurring). We also provide some examples of it's use in rendering algorithms frequently used in graphics.

Interactive Level-of-Detail Selection Using Image-Based Quality Metric for Large Volume Visualization

“…The metric makes use of multiresolution wavelet decompositions of the query and database images, and operates on the coefficients of these decompositions. Gaddipati et al [7] presented a wavelet-based metric which captures the change in images wrought by operators and the image synthesis algorithms. Sahasrabudhe et al [27] proposed a quantitative technique which accentuates differences in images and data sets through a collection of partial metrics.…”

“…These metrics have clear physical meanings and are simple to compute. However, they are usually not effective in predicting the quality of the rendered images due to the lack of correlation between data and image, as indicated in [7], [15], [22], [27], [30]. Image-based metrics focus on the ultimate images the user perceives, and strive to capture the quality loss in the rendered images introduced by rendering low resolution data.…”

Structured spatial domain image and data comparison metrics

“…The earliest metrics: mean square error (MSE), root mean square area (RMSE), and peak signal-to-noise ratio are spatial domain metrics. However, these metrics often arrive at a global measure that can sometimes overstate the distortion a user perceives in the image [3]. This lack of correlation between the type of error in an image and the response of the human visual system to varying types of errors led researchers to begin developing image metrics.…”

“…After transforming and weighting by a HVS model, the metric computes the MSE between the two images as an overall quality measure. A variety of approaches using wavelets in various capacities as image metrics have been reported [3] [5]. Gaddipatti et al employed a metric to guide the visualization process [3].…”

“…In response, we propose structured spatial metrics in this paper. Numerical comparison metrics have been proposed elsewhere [3] [15]. The proposed metrics differ from previous ones, in that, they operate mainly in the spatial domain and attempt to reveal structured differences.…”