Low Intrusive In-image Text Overlay Based on Saliency Cut and Computational Aesthetics

• Main Authors: Mien-Tsung Tsai, 蔡綿綜
• Other Authors: Jen-Shin Hong
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/33219226581006567627

博士 === 國立暨南國際大學 === 資訊工程學系 === 102 === Visual aesthetics is an important aspect in user interface and product design (Liu, 2003). Specifically for the design of computer interfaces, much research has recently been devoted to investigate the impact of interface aesthetics on the overall user experiences when interacting with a computer system. Although the exact mechanisms that link aesthetics and system usability are not yet fully clear, the importance of aesthetics in system interface design has certainly been more clearly demonstrated and received increased attention. This thesis considers the aesthetics of in-image text overlay which consists of a large-size background image with a paragraph of texts overlaid on it. Applications using text-overlaid image interface typically are graphic-intensive multimedia applications that attempt to use the background image to generate appropriate aesthetic effects. This thesis aims to quantitatively analyze the visual aesthetics of text-overlaid image such that an optimal way to partition and position the overlaying texts can be computed automatically. We first developed computational models for white-space fractions and visual balances for text-overlaid images. The relationship between the white-space and the aesthetic appeal was experimentally investigated. A robust technique adapting from saliency cut techniques for detecting the background region of an image on which the texts should be overlaid was developed. Based on these results, we have implemented an intelligent system for composing text-overlaid images. The system automatically partitions the input texts to a sequence of paragraphs and then chooses the optimal position in a given photo for each partitioned paragraph such that the final white-space fraction and average visual balance is optimized. The results obtained by our prototype system show a promising future.