Digital Imaging and Halftoning Algorithms Design: Parallel Processing, Printed Image Artifacts Modeling, and Learning-Based Image Analysis

This thesis explores four topics in designing imaging processing algorithms. The first one is enhancing computation efficiency and implementation flexibility. We propose a parallel processing architecture for the error diffusion algorithm, allowing it to process multiple rows of image pixels simultaneously. We also develop a new loss function in the training system to minimize the weighted mean squared error in the power spectra. With this new algorithm, better halftone reproductions and higher hardware efficiency can be achieved. The second topic is improving the image quality of printed images. Two novel printer models for the direct binary search (DBS) algorithm are proposed. One is called the dot profile model. It is used to compensate dot shape irregularity errors of inkjet printers. The other one is called the ink drop displacement model. It is intended primarily to compensate for the stochastic ink drop displacement error in the prints. For both models, we first characterize the statistical properties of the printed dots and then integrate this prior knowledge in the fidelity metric which directs the image reproduction of DBS. In so doing, the image quality is greatly enhanced. The third topic is supervised learning algorithms applied to document image analysis. In this application, we develop a set of features to use with a support vector machine model to discriminate document images based on the color information and contents. The last topic is learning-based makeup shade matching. We train machine learning models to predict the best matching foundation shade based on the selfie images of a customer. Besides, a novel color correction algorithm is designed to minimize color inconsistency.