The Research Of Image Line Drawings Based On Conditional Random Fields

With the rapid development of the imaging technologies, including the medical imaging and multi-band military remote sensing imagery, with the widespread adoption of this digital technology for civilian, such as the digital cameras, camcorders and web cameras etc, to get the images and video information becomes much more convenient than ever. However, the traditional manual processing has apparently lagged behind the explosive growth of the amount of images and video information, therefore, a good image processing method is urgent. In this paper, the theory of the image line drawing is introduced to resolve the problem of the image and video information quickly understood by the computer. Based on the existing detection methods of the image edges, borders and contour, the image borders detection method based on conditional random fields model and a line drawings generation from images algorithm based on edge tracking are given. The contents of this research are as follows:(1) The image borders detection method based on conditional random model is divided into three phrases:data preprocessing, model training and inferring, inferred data post-processing. In the first phrase, Preprocessing includes the image feature calculation and post-processing. In the image feature calculation, we calculate the oriented energy, gray gradient, texture gradient and statistical feature. In the post-processing features, we use the median filter algorithm to solve the double edge problems of the image texture gradient. Furthermore, we get a combined gradient by combining the gray gradient and the texture gradient. Finally, we thin these three features of image oriented energy, image statistical feature and combined gradient with the non-maxima suppression. During the second phrase, training and inference are included. The training data pair obtained with the low frequency after wavelet translating of the original image. Moreover, we use the vector group which are consists of oriented energy, statistical feature and combined gradient as observational data, and use the learning boundary image as the tag data, which is a image after the texture edge, noise and background edge suppression based on that vector grope. In the inference, Gibbs sampling algorithm is also used in this phrase. In the last phrase, we present a method to detect, refine and connect the boundary points, which deal with the result of inferred data. Consequently the experimental analysis is given in the paper, and the results show that our method is feasible and effective.(2) The line drawing images algorithm includes edge tracking and line drawing. In this paper, an edge tracking algorithm based on dissimilarity measure is presented in order that the detected borders can be classified and connected. Meanwhile, we employ non-uniform B-spline to interpolate the discontinuous borders, and introduce Gaussian function to obtain a smooth boundary line. Subsequently the brush corresponded with the curvature of the line is generated, and the line drawings of images is achieved ultimately. Experiment results show that our method is able to quickly generate higher-quality line drawings.