Laparoscopic Image Recovery and Stereo Matching
Abstract
Laparoscopic imaging can play a significant role in the minimally invasive surgical procedure. However, laparoscopic images often suffer from insufficient and irregular light sources, specular highlight surfaces, and a lack of depth information. These problems can negatively influence the surgeons during surgery, and lead to erroneous visual tracking and potential surgical risks. Thus, developing effective image-processing algorithms for laparoscopic vision recovery and stereo matching is of significant importance. Most related algorithms are effective on nature images, but less effective on laparoscopic images.
The first purpose of this thesis is to restore low-light laparoscopic vision, where an effective image enhancement method is proposed by identifying different illumination regions and designing the enhancement criteria for desired image quality. This method can enhance the low-light region by reducing noise amplification during the enhancement process. In addition, this thesis also proposes a simplified Retinex optimization method for non-uniform illumination enhancement. By integrating the prior information of the illumination and reflectance into the optimization process, this method can significantly enhance the dark region while preserving naturalness, texture details, and image structures. Moreover, due to the replacement of the total variation term with two $l_2$-norm terms, the proposed algorithm has a significant computational advantage.
Second, a global optimization method for specular highlight removal from a single laparoscopic image is proposed. This method consists of a modified dichromatic reflection model and a novel diffuse chromaticity estimation technique. Due to utilizing the limited color variation of the laparoscopic image, the estimated diffuse chromaticity can approximate the true diffuse chromaticity, which allows us to effectively remove the specular highlight with texture detail preservation.
Third, a robust edge-preserving stereo matching method is proposed, based on sparse feature matching, left and right illumination equalization, and refined disparity optimization processes. The sparse feature matching and illumination equalization techniques can provide a good disparity map initialization so that our refined disparity optimization can quickly obtain an accurate disparity map. This approach is particularly promising on surgical tool edges, smooth soft tissues, and surfaces with strong specular highlight.