|作品名稱（其他語言）||On Professional Contemporary Style Photographing Instruction Based on Neural Tree Based Classifiers Applied to Image Aesthetics Assessment|
|摘要||In this dissertation, we study on how to use artificial intelligence and data mining technologies to make computers able to perceive the concept of beauty, which is an abstract idea, and design a photographing instruction system accordingly. We collect contemporary style images captured in recent years on social networks for analysis. In our instruction system, there are two parts of instruction, one is image characteristics, and the other is image composition. The image characteristics refers to the color and textures, while the image composition refers to the structure of an image.
Our proposed photographing instructor is composed of tree-based classifiers and artificial neural networks, and form a random forest to predict whether an image meets the criterions of the contemporary style. Binary decision tree are built for photographing instruction. However, the decision tree suffers from axis-aligned problem, which limits its accuracy. Therefore, we combine the decision tree and neural network, and use the subsets to build multiple random trees as random forest to improve the accuracy. We also described about the limitations of the instruction system. The system gives semantic sentences to users for image characteristics enhancement, and use blocks to indicate which regions should be improved for image composition.
In the experiments, we predict whether an image is favorable. When using image characteristics and composition features separately, and achieved 85% accuracy. When combining the two types of features, the accuracy was above 91%. In addition, the proposed instruction system is able to give correct suggestions. After applying the suggestions from our proposed system, the colors were more harmonized, the compositions were more balanced, and the main subjects were enhanced.
|關鍵字||計量審美學;資料探勘;決策樹;隨機森林;類神經網路;Computational aesthetics;data mining;decision tree;random forest;artificial neural networks|