Imaging Systems Research: Digital Color Halftoning, Color Management, and Person Segmentation
The efforts of this dissertation are divided into three aspects: clustered-dot digital halftoning, color management and person segmentation. The requirement for high-quality print products grows rapidly alongside the development of technology, artwork, and commercial, etc. And halftoning algorithms are the methods to manage the deposit of colorants, which determine the printed image quality directly. However, the conventional color halftoning method for laser electrophotographic printers is based on periodic clustered-dot halftoning screens, which will result in visible moire ́ and rosette. In order to overcome these artifacts coursed by the interaction of periodic screens, three color aperiodic clustered-dot halftoning algorithms, PARAWACS-CLU-DBS, NPAC-CLU-DBS with concentric-ring structure and NPAC-CLU-DBS with separate-cluster structure, are proposed. Because of the aperiodic cluster structure, all of the three algorithms have an inherent stability during the printing process and reduce the periodic artifacts dramatically. Here NPAC stands for Neugebauer Area Primary Coverage. Additionally, as the halftoning algorithms based on NPAC, all of the three new algorithms require NPAC images as input. So that, in the second part of this dissertation, a color management pipeline for NPAC-CLU-DBS halftoning algorithms is proposed. Previously, the source (sRGB) color gamut is mapped directly into the destination (printer) color gamut. As a result, a large color mismatch appeared in the printed image because the source color gamut is much larger than the destination color gamut. In this section, a new color management method with image-dependent color gamut mapping is proposed, which can make the most use of the printer color gamut so that to reduce the color mismatch between the continuous-tone image and the printed halftone. In the third part of this dissertation, a novel automatic person segmentation system is presented. Automatic person segmentation has many applications in the real- world, but both the higher segmentation precision requirement and large variation of poses and dressing styles increase its challenge. To tackle these problems, we first proposed a light weight automatic person segmentation neural network (LWPSN) in this part. With a smaller model size and higher segmentation accuracy, our LWPSN network has the potential to be applied to mobile devices. Second, in our person segmentation system, it is the first time to incorporate the pose estimation to improve the segmentation performance. A novel spatial saliency map is generated based on the pose estimation to provide more spatial information so that to improve the performance of our segmentation network. Additionally, during the data augmentation process, we also proposed the image resizing augmentation method for person segmentation, which can push the network learning spatial information more efficiently. With all these efforts, our person segmentation model outperforms the state-of-the-art method on accuracy for the person segmentation task. Furthermore, we also try to reach out the automatic portrait matting method in order to achieve better foreground transition result.
Lin, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our