This paper proposes a no-reference image quality evaluation model for accurately assessing the quality of real-world images displayed on head-mounted display (HMD) devices. The proposed model employs a simulation of human visual system, providing a reliable measure of image quality. Initially, an efficient convolutional neural network (CNN), specifically designed for noise characteristics, is utilized to obtain a near-perfectly noise-reduced image. The difference between this image and the target image is then calculated in the linear domain. To emulate the contrast sensitivity and masking effects inherent in the human visual system, we introduce a sophisticated frequency-domain filter model in a uniform color space. The resulting multidimensional data from the filters are aggregated and corrected based on the average brightness. Our model's performance is validated against Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics using the TID2013 dataset, revealing superior correlation coefficients. Human factors experiments further confirm the model's reliability and practicality in real-world scenarios.
|