紋理合成貼圖的注視點與評分預測

Abstract

有鑑於目前的研究，大多是利用數學方程式去對紋理貼圖合成的結果做評分，這些評分的標準並不適用在各式各樣的紋理貼圖演算法，而且這些評分的預測也不一定完全符合人類的評分標準，所以我們提出一個有效的"特徵粹取"並搭配"眼動儀資料"訓練一個評分模型。

相較於非感知的評分方式，我們的model因為結合了人類感知的資料，所以複雜了許多，因此我們將實驗的model分成兩個階段，第一個是虛擬注視點 (visual attention)，第二個是人類感知評分預測 (perceptual rating)，在結果階段我們會比較有無注視點是否會對結果造成好或壞的影響來證明我們這兩階段的必要性。

Texture synthesis is a hot topic in computer graphics; however, there is less work on perceptual evaluation of synthetic structural texture. As visual attention is the first stage of visual cognition process, we propose two models, visual attention model and perceptual rating model, to predict visual saliency and human rating on synthetic structural textures. We designed an experiment to gather subjects' eye-tracking data and rating while evaluating the similarity of synthesized textures. The visual attention model is developed to associate texture features and fixations. The perceptual rating model is trained to associate the relationship between the fixations and the rating. We compared our visual attention model with the saliency map. Our model correctly predicts 82.7% of fixation positions while the saliency map only achieves 57%. For the perceptual rating, Chi-square value of our model is 3.98 but non-perceptual metric is 6.95, comparing to human's rating scores. Our model is very helpful for guiding texture synthesis and manipulation algorithms to efficiently allocate computational resources to those regions that humans may consider unnatural and pay attention to.