Three problems in image analysis and rendering: Local defect detection, semantically-based 2.5D printing, and aesthetic evaluation of fashion photos
In this dissertation we will discuss three problems in image analysis and rendering. The main topic is the aesthetic evaluation of fashion photos. We develop an autonomous online fashion shopping photo aesthetic quality predictor with support vector machine (SVM) algorithm. In this project, we design multiple image features to represent photo aesthetic quality. In addition, we adopt the elastic net method to rank the features based on their relevance with the photo aesthetic quality. In order to yield the best prediction accuracy, the wrapper feature selection method is used to produce the optimal training feature subset. Besides the SVM predictor, we incorporate the deep neural network technique to train another predictor and compare the accuracy with the SVM predictor. Finally, we propose a Bayesian approach to infer photo aesthetic quality scores from psychophysical experiment, and prove that with small number of ratings our method gives more robust estimation of the true aesthetic quality score compared with traditional approaches such as rounded average score. For the other topics included in this dissertation, I will first describe an algorithm for the detection of a specific class of print quality artifacts: local defects, and the prediction of the overall test page quality that would be assigned by an expert observer. Subsequently, I will introduce a method to reproduce a given textured area of an image with a novel printing technology, namely 2.5D printing, based on the semantic information of that area. 2.5D printing is capable of representing texture in a more appealing way to the human observer due to the fact that it can reproduce the tactile details and structures of a texture.
Allebach, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our