THREE PROBLEMS IN IMAGE ANALYSIS: IDENTIFICATION OF GRAY SPOT DEFECTS IN PRINTS FROM ELECTROPHOTOGRAPHIC PRINTERS, COLOR PREDICTION OF INTERACTION FOR INK-JET PRINTERS, AND FACE RELIGHTING

Printing quality is very important. For many years, many researchers contribute to achieve appealing printed objects by reducing printing defects and visual noises. However, manually detecting the printing defects is very time consuming, which then requires a more efficient way to automatically detect defects. Besides, a lot of models have been built to investigate human visual response to printing noise in macro level. Few of them investigate the relation between printing quality and micro-level structures of ink patterns. In addition to directly improve printing quality, researchers also build models to simulate printing process, which, in turn, will help improving printing system design. But, due to the complexity in printing systems, it is very hard to use traditional models to simulate the printing process.

In our research, for electrophotographic printing, we design a segmentation-based framework to automatically detect printing defects in scanned electrophotographic printed pages. For ink-jet printing, we build a color separation framework to segment different color components of printed pages on different coated papers with different ink-drop sizes. And based on our segmentation results, we also investigate the relation between paper types, ink drop sizes and grain image noise in composite color printing. Besides, based on our previous research about ink-jet printing at the microscopic level, we leverage deep learning methods to simulate the printing process at the microscopic level. Comparing with traditional printing simulation methods, our simulation model is an end-to-end framework working on color printing at very high resolution.

The second topic focus on image enhancement to do face relighting. Poorly exposed images degrade the performances of models for different tasks, such as face recognition, face landmark detection and skin-tone detection. Enhancing the photographs to generate balanced lighting distributions over faces will help to increase the performances of models. Previously, a lot of research has been done to tackle this problem and has made great progress. One kind of popular method in poor exposure correction is the ratio imaging based method. It separates reflectance from lighting distribution. A lot of methods in this class estimate the lighting distribution from the pixel level, which includes a lot of textures from reflectance. This results in low contrast enhanced images. They ignore the 3D face geometry. Also it is difficult to deduce the 3D face normal from a single view. But human faces have regular shapes so we can use a prior 3D morphable model to fit the target face. And also, based on the development in face alignment, it becomes possible to acquire the face geometry from a single view in higher accuracy. Based on the development in face alignment and ratio imaging based methods, we leverage the 3D face geometry information to better enhance poorly exposed face images. We first estimate the lighting distribution accurately. Then we use an optimization process to further refine the estimated lighting distribution. This two-step method generates good results with balanced lighting distribution.