A Study of Image Moment-Based Painterly Rendering Algorithms

• Main Authors: Lung-Hung Yang, 楊龍鴻
• Other Authors: Chung-Ming Wang
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/57365728190367740166

碩士 === 國立中興大學 === 資訊科學研究所 === 93 === Painterly rendering has recently been under intensive studies in computer graphics communities. Painterly rendering algorithms based on the image moment have demonstrated some success, producing visually plausible painting images. However, most of these algorithms employed line-based brush strokes, disregarding the curved-strokes natures that have been adopted by painters in the real world. In this thesis, we present two novel curved-strokes algorithms based on the image moment for painterly rendering. We present the first algorithm: A Novel Moment Based Painterly Rendering Algorithm with Curved Brush Strokes. Our algorithm employs the line integral convolution technique to determine a curved line passing thorough pixels coincident with the gradients of the image. Then, along with visiting every pixel in the line, we calculate the magnitude of the image moment to determine whether a stroke is to be drawn on the pixel currently visited. Next, we determine the sizes and directions of the stroke automatically. We first compare the variances of the image moment at the current pixel, with those of the image moment in the neighborhood pixels. We then determine the suitable magnitude of sizes and directions through a linear interpolation calculation. Under this mechanism, a stroke with a small size and short length will be assigned to the pixel with small magnitude of the image moment. Moreover, we apply an image segmentation technique to identify boundaries of objects. When proceeding to painterly rendering, we can make sure that strokes will all be maintained inside the objects, without passing through the boundaries. Finally, we use a Gaussian filter on the original image and replace those pixels that are not drawn by any strokes. This allows us to simulate painterly rendering images when considering the existence of the canvas. To the best of our knowledge, our algorithm is novel for image moment based painterly rendering with curved strokes. Due to the curved and local strokes, our algorithm generates painterly rendering images that are similar to the stroke styles painted by the artists. We propose the second algorithm: A Moment Based Painterly Rendering Algorithm for High Dynamic Range Painting. Here, we extend our rendering research in the Low Dynamic Range Image and propose a novel algorithm for the High Dynamic Range image. First, our algorithm classifies the luminance of an HDR image into four different regions, producing a pixel luminance label for the HDR image. Then, the algorithm employs the line integral convolution method to construct the curved brush strokes. Meanwhile, we reproduce the HDR image to the LDR image using a tone mapping technique. This LDR image is then segmented into different regions, producing a pixel color region label. By referring to both the luminance label and color region label, we then render the HDR image with painterly style, ensuring that the strokes will remain inside the boundaries of an object. This guarantees the visual fidelity of the painterly rendering HDR image. Finally, our algorithm adopts the Gaussian filter to simulate the effects of canvas and covers those pixels that have no strokes on them. To the best of our knowledge our algorithm is the first to cope with the HDR image for painterly rendering. Combined with the tone mapping techniques, our algorithm is able to produce many different styles of painterly rendering images. In addition, our program successfully developed the curved strokes for the HDR images, so the resultant image retains the characteristics of curved stroke, making it to faithfully simulate the painting styles of the artists. In this thesis, we propose two novel image moment-based painterly rendering algorithms. Our algorithms preserve the curved strokes, and they are suitable for LDR and HDR images. Experimental results demonstrate these two algorithms generate painterly images that are with satisfactory visually plausible appearance.