We developed a neural network-based method for evaluation of display luminance and color non-uniformity (which we call Mura). We studied a correlation between our developed method and human visual evaluation because visual evaluation is the gold standard for Mura evaluation. We achieved Pearson correlation coefficient of 0.82.

We developed a method for automated evaluation of display luminance non-uniformity using an auto-encoder. Usually, a reconstruction loss of auto-encoder is used for abnormality detection. In our method, we used reconstruction loss as the main indicator and cosine similarity as a secondary indicator. Our method succeeded in the non-uniformity evaluation.

We developed a simplified tool for measuring image quality of medical liquid-crystal displays (LCDs) using a commercially available color digital camera. This tool implemented as a plug-in software for ImageJ (open-source image processing program) was designed to compute modulation transfer functions (MTFs) and Wiener spectra (WS) of monochrome and color LCDs from LCD photographed images captured by a camera. The intensities of the red (R), green (G), and blue (B) signals of the unprocessed image data depend on the spectral sensitivity of the image sensor used in the camera. In order to evaluate image quality based on LCD luminance, the plug-in software calibrates the RGB signals from the camera using measured luminance of the LCD and converts them into grayscale signals that correspond to the luminance of the LCD. The MTFs and WS are determined based on the line response from a one-pixel line image and the one-dimensional noise profiles acquired by scanning the uniform image using numerically synthesized slit, respectively. With this plug-in software for ImageJ, we are able to readily compute MTFs and WS of both monochrome and color LCDs from unprocessed image data of cameras. Our simplified tool is helpful to evaluate and understand the physical performance of LCDs for a large number of display users in hospitals and medical centers.

No abstract

We developed a quantitative method for evaluation of display luminance and color non‐uniformity (which we call mura) using a deep convolutional neural network (DCNN). In previous research, quantitative evaluation methods using DCNN were studied. From among a wide variety of DCNN models, these methods have used a convolutional autoencoder (CAE) for abnormality detection. A CAE trained with only non‐defective data was then used to evaluate the degree of mura. However, there is a problem with these methods in that they have not been able to evaluate properly when there are multiple mura defects on a single screen. Consequently, the correlation between these methods and human visual evaluation was low. Therefore, we improved the previous method to handle multiple mura defects on the same screen. We confirmed the accuracy of the proposed method by comparing the quantitative evaluation and human visual evaluation. As a result, we obtained a higher correlation than the previous methods.