Integration of human color vision models into high quality image compression

Strong academic and commercial interest in image compression has resulted in a number of sophisticated compression techniques. Some of these techniques have evolved into international standards such as JPEG. However, the widespread success of JPEG has slowed the rate of innovation in such standards. Even most recent techniques, such as those proposed in the JPEG2000 standard, do not show significantly improved compression performance; rather they increase the bitstream functionality. Nevertheless, the manifold of multimedia applications demands for further improvements in compression quality. The problem of stagnating compression quality can be overcome by exploiting the limitations of the human visual system (HVS) for compression purposes. To do so, commonly used distortion metrics such as mean-square error (MSE) are replaced by an HVS-model-based quality metric. Thus, the “visual” quality is optimized. Due to the tremendous complexity of the physiological structures involved in the human vision task, HVS-models are rather based on psychophysical observations. Physiologists and psychologists have done psycho-visual experiments with the goal of understanding how the HVS works. However, the experimental conditions used are not representative of many image processing applications. Nevertheless, engineers apply the results of those psychovisual experiments in these applications. To do so, they use the simplified HVS-models with little knowledge regarding the applicability of these models under the new conditions. This thesis addresses this problem by designing new psycho-visual experiments and models designed explicitly for direct incorporation into image compression applications. One commonly mentioned limitation of the HVS concerns the reduced sensitivity for patterns of high spatial-frequencies. Exploiting this behavior can significantly improve the compression quality. The phenomenon is parameterized by the contrast sensitivity function (CSF). Even if the sensitivity is a function of spatial frequency and color, the CSF is typically only implemented for the luminance channel in color images. Here, a proper CSF-modeling for color is presented. It implies new color CSF measurements and a new method to incorporate the CSF with significantly improved precision into compression schemes. Furthermore, a consequent analysis of the final performance for three different CSF-enhanced opponent color spaces is presented. The sensitivity of the HVS for compression artifacts also varies with respect to the strength of local contrasts. Thus, an artifact might be hidden by the presence of strong contrasts or locally active image regions. This phenomenon, referred to as masking, was measured and modeled in numerous experiments using sinusoidal patterns and noise stimuli. With respect to image processing applications, those test patterns are over-simplified. Therefore a new method is proposed that uses photo-realistic images as stimuli for masking experiments. It is used to evaluate and predict the performance of various masking models that are popular candidates for compression applications. This comparative analysis also supports an improved parameterization of masking effects. Subjective quality evaluations showed that the HVS is very sensitive to the loss of texture information, because blurred images appear unnatural. However, the exact encoding of texture information is bitrate intensive. To avoid this, the newly proposed scheme uses texture-modeling. The texture is characterized by only a few parameters that can be encoded for a modest increase in the bitrate. Even if the synthesized texture is pixel-wise different from the original, the HVS is nevertheless deceived due to their apparent similarity. All newly presented techniques can be incorporated into the compression standard JPEG2000. Finally, for a given subjective image quality, the HVS-enhanced JPEG2000 consumes less than half the bitrate needed with the “default” JPEG2000 scheme.