Development of a Perceptually Calibrated Objective Metric for Auto White Balance
Elaine W. Jin,Yixuan Wang,Wentao Liu
DOI: https://doi.org/10.2352/issn.2470-1173.2018.12.iqsp-342
2018-01-01
Electronic Imaging
Abstract:This study aims at developing an image quality metric for camera auto white balance (AWB), with a transform to just noticeable differences (JNDs) of quality in pictorial scenes. In this study, a simulation pipeline was developed for a Nikon D40 DSLR camera, from raw capture to rendered image for display. Seven real-world scenes were used in the study, representing capture conditions in outdoor daylight, indoor fluorescent lighting, and indoor incandescent lighting conditions. Two psychophysical experiments were performed, and 38 observers participated in the study. In study one, method of adjustment was used to explore the color aims for individual scenes. In study two, a softcopy quality ruler method was used to refine the color aims and define the quality falloff functions. A quartic function was used to fit the results from the softcopy ruler study, forming the proposed objective metric for camera auto white balance. Introduction The impact of digital photography has been growing due to the ubiquitous presence of the smartphones. The cameras equipped on these smartphones have seen steady improvement in quality, and today the quality level has reached and even surpassed that of the compact digital still cameras. To continue the improvement on camera image quality, it is essential to develop objective metrics for image quality that can correlate with human perception. Image quality attributes can be divided into two major groups, artifactual attributes and preferential attributes [1]. Artifactual attributes refer to defects introduced by imaging systems, including attributes such as noise, blur, and color shading. For these attributes the ideal aim positions are well defined, i.e. free of the defects. Preferential attributes refer to attributes that are preferential in nature, including all color/tone rendering related attributes such as color saturation, contrast, exposure, and white balance. It is challenging to define the ideal aim positions for such attributes because the aim positions are likely observer and scene dependent, Auto white balance (AWB) is a camera auto function that detects the scene illumination and compensates for it in image rendering. In an ideal scenario, the camera image processing pipeline can perfectly detect the scene illumination, and rendering a neutral patch to be perfectly neutral in the given output color space (e.g., the sRGB color space) [2]. In such a scenario the measurement of the quality of the camera AWB function becomes somewhat straightforward. For example, in the Imatest Master software tool the white balance error is defined as the average CIELAB chroma of patches 20 – 23 on an X-Rite ColorChecker Classic target [3] [4]. There are complications to this ideal scenario. It has been reported that the preferred color/tone reproduction in photographic images are different from the colorimetric reproduction [5-6]. For example, people prefer a blue-sky image to have much higher purity than the real sky blue. Today’s smartphone cameras also adjust the white balance rendering aims according to the scene illumination. Fig. 1 shows the images of an X-Rite ColorChecker Classic target captured by two smartphone cameras under two lighting conditions. The two daylight images looked very similar to each other, and the neutral patches are close to neutral. In comparison, the images captured under tungsten light are rendered warmer, with the neutral patches moving towards the red/yellow direction. Furthermore, it is interesting to notice that these two images looked very different from each other, with one having much higher chroma than the other. These observations inevitably lead to the following questions: what are the preferred white balance aims for images captured under different lighting conditions? What is the quality falloff when the rendering positions deviate from the preferred color aims? The purpose of this study is to answer these two questions.