Browsing by Author "Cao, Xingdong"

Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Image-to-Image Translation with Application to Biometrics
    (2022-01-27) Cao, Xingdong; Yanushkevich, Svetlana; Smith, Mike; Uddin, Gias; Goldsmith, Peter
    This thesis focuses on two image-to-image translation tasks related to the field of face biometrics. We use conditional generative adversarial networks (cGAN) with different loss combinations to solve these two tasks. The first one addressed the application, such as face detection in the images taken in challenging scenes with high variations of light known as wide dynamic range images. To improve the imaging, high dynamic range (HDR) cameras that offer near-perfect exposure in both bright and dark areas simultaneously are used instead of regular cameras. This research work offers an approach to converting HDR images into low dynamic range (LDR) images, called High-to-Low (H2L) conversion. For the H2L task, HDR images with various contents were collected, and the ground-truth LDR images were generated by 30 different tone-mapping operators (TMOs). The quality of images was assessed using the tone-mapped image quality index (TMQI). We used a cGAN loss, perceptual loss, and feature matching loss to train the cGAN and utilized TMQI to evaluate the generated LDR images. We also tried different training and testing schemes to reach a better visual effect and higher TMQI. The TMQI of the generated LDR images in the testing database consisting of 105 images reaches 0.90, which outperforms all other contemporary tone mapping operators. An extensive face detection technique was applied to detect faces on the generated LDR images, to reach the high accuracy of such detection. The second task addressed face biometrics in the infrared (thermal) spectrum. We consider converting a thermal face image into another thermal face image given a target face temperature, that is, Thermal-to-Thermal (T2T) conversion. To the best known of our knowledge, it is the first time that such conversion was attempted. Images and thermal information from Carl Database were used to train a temperature predictor. We used a cGAN loss, perceptual loss, and temperature loss to train the cGAN. The state-of-the-art face recognition techniques were utilized to test the generated thermal face images belonging to the same person, given different face temperatures. The generated thermal images reached a rank-1 face recognition accuracy of 91 %.