Convolutional neural network for distortion Classification in face images.

Abstract

Face processing algorithms are becoming more popular in recent days due to the great domain of application in which they can be used. As a consequence, research about the quality of face images is also increasing. The current approach to Face Image Quality Assessment (FIQA) is focused on improving the performance of face recognition systems, as a result, current FIQA algorithms don't provide an indication of quality, but a performance estimation for face recognition algorithms. This approach makes the FIQA algorithms potentially unsuited for other scenarios regarding face images, and susceptible to inherit the limitations of face recognition. The present work tackles the main limitations of the current FIQA algorithms by proposing a new approach based on the distortions affecting the images. We developed two models based on Convolutional Neural Networks (CNN), to classify facial images according to the type and the degree of the distortion present in them. The models' output provides qualitative information about the quality of facial images, useful for face recognition systems, as well as other face processing algorithms. Additionally, the proposed method can be a starting point to image enhancement processes like denoising, and deblurring. Two other contributions can be outlined from this work: a comprehensive study about the impact of blur, noise, brightness, contrast, and JPEG compression in face processing algorithms; and a new dataset for image quality assessment and distortion classification in face images.

P‐168: Threshold Estimation Experiment of color difference between on‐axis and off‐axis using various complex images

“…Therefore, viewing angle characteristics of images experienced by users have an influence on image quality, and several studies have been conducted the image quality depending on the viewing angles. [1,2] To measure the viewing angle performance of the display, however, white patches or simple color patches have been mainly used instead of images. These patch-based measurement method cannot take into account what the user actually experienced, as the color of the patch may differ from that of the image.…”

Radiomics and Artificial Intelligence Study of Masseter Muscle Segmentation in Patients With Hemifacial Microsomia

“…11 The necessary training data and segmentation gold standard is usually obtained by manual segmentations. [12][13][14][15] Kohl et al 16 proposed U-Net neural convolution network for medical image segmentation, recognizing features and position of the region of interest, and extended the U-Net from a plane to a 3-dimensional (D) structure. After adaptive training for the masseter muscle and mandible, we proposed a deep learning model based on U-Net for multiorganization segmentation.…”