Image based human body rendering via regression & MRF energy minimization

Abstract

A machine learning method for synthesising human images is explored to create new images without relying on 3D modelling. Machine learning allows the creation of new images through prediction from existing data based on the use of training images. In the present study, image synthesis is performed at two levels: contour and pixel. A class of learning-based methods is formulated to create object contours from the training image for the synthetic image that allow pixel synthesis within the contours in the second level. The methods rely on applying robust object descriptions, dynamic learning models after appropriate motion segmentation, and machine learning-based frameworks.

Image-based human image synthesis using machine learning is a research focus that has recently gained considerable attention in the field of computer graphics. It makes use of techniques from image/motion analysis in computer vision. The problem lies in the estimation of methods for image-based object configuration (i.e. segmentation, contour outline). Using the results of these analysis methods as bases, the research adopts the machine learning approach, in which human images are synthesised by executing the synthesis of contour and pixels through the learning from training image.

Firstly, thesis shows how an accurate silhouette is distilled using developed background subtraction for accuracy and efficiency. The traditional vector machine approach is used to avoid ambiguities within the regression process. Images can be represented as a class of accurate and efficient vectors for single images as well as sequences. Secondly, the framework is explored using a unique view of machine learning methods, i.e., support vector regression (SVR), to obtain the convergence result of vectors for contour allocation. The changing relationship between the synthetic image and the training image is expressed as a vector and represented in functions. Finally, a pixel synthesis is performed based on belief propagation.

This thesis proposes a novel image-based rendering method for colour image synthesis using SVR and belief propagation for generalisation to enable the prediction of contour and colour information from input colour images. The methods rely on using appropriately defined and robust input colour images, optimising the input contour images within a sparse SVR framework. Firstly, the thesis shows how contour can effectively and efficiently be predicted from small numbers of input contour images. In addition, the thesis exploits the sparse properties of SVR efficiency, and makes use of SVR to estimate regression function. The image-based rendering method employed in this study enables contour synthesis for the prediction of small numbers of input source images. This procedure avoids the use of complex models and geometry information. Secondly, the method used for human body contour colouring is extended to define eight differently connected pixels, and construct a link distance field via the belief propagation method. The link distance, which acts as the message in propagation, is transformed by improving the low-envelope method in fast distance transform. Finally, the methodology is tested by considering human facial and human body clothing information. The accuracy of the test results for the human body model confirms the efficiency of the proposed method.