- 學位論文
探討影像顯示大小如何影響美感
Investigation of How Image Display Size Affects Image Aesthetic Perception
摘要
影像美感評估是近幾年非常熱門的研究議題，之前相關的研究指出影像對比、顯著的特徵(saliency features)、和物體在影像中的位置會交互影響一張影像看起來的美學觀感。然而之前的研究沒有探討到影像所顯示的大小如何影響美學觀感，因此我們探討影像顯示大小與影像美學觀感兩者之間的關係。 本論文中，我們探討影像的解析度(pixels)和實際大小(physical dimension)如何影響影像的美感。我們利用群眾外包(crowdsourcing)技術蒐集到100張影像，及其分別顯示成不同大小的美學分數，並且證實了影像顯示大小的確會對美學觀感造成複雜的影響。整體來說愈大張的影像有比較好的美學觀感，但是一張影像看起來是否好看還是取決於它的內容。我們取出與內容和配置有關的特徵(content and compositional features)，並利用偏最小平方差迴歸(PLSR)建立分數模型(scoring model)和排名模型(ranking model)預估影像的美學分數和相對的排名。我們藉由分析每個特徵在預估美學分數的效能和訓練出的權重來發現那些特徵對於評估影像美學觀感比較有效。我們討論許多議題，例如不同種類影像各自的效能、使用多少比例的資料去訓練如何影響系統的預估效能、和關於影像是否有最佳的顯示大小等。本論文提出當我們要評估一張影像的美學觀感時，我們必須將影像的內容以及影像所顯示的大小考慮進去，希望本篇論文能夠為之後關於影像美學觀感方面的研究提出一些啟發。
並列摘要
Image quality assessment has become a popular topic. Previous related works have shown that image contrast, saliency features, and composition of objects may jointly decide whether an image looks good or not. However, how image display size affects image aesthetic perception has not been investigated in previous works. We conduct a pilot study to verify the effect of display size on image aesthetic perception. In this thesis, we devote to investigate how an image’s resolution (pixels) and physical dimension (inches) affect how much viewers appreciate an image. We use crowdsourcing technology to collect a large-scale aesthetic assessments of 100 images displayed in a variety of physical dimensions, and show that an image’s display size significantly affects its aesthetic rating in a complicated way; normally an image looks better with a bigger display size, but it may look relatively worse depending on its content. We use partial least square regression (PLSR) to develop the scoring model and the ranking model to predict an image’s aesthetic rating and relative aesthetic ranking based on its content and compositional features. We discover effective features in aesthetic prediction by analyzing the relationship between performance and learnt weights. We discuss performance for different categories and how amounts of training data affect prediction performance. We further discuss an interesting issue about the best display size of each image to inspire future research. We hope that this work will shed some light on future research by revealing that both content and presentation should be considered in image aesthetic evaluation.