Assessing photograph aesthetic quality with color based descriptor
Xianghui Zhu, Teng-Sheng Moh · 2016
Assessing photograph aesthetic quality is an interesting and important problem of computer vision. The understanding of how to classify photographs based on their aesthetic value not only provides us with important knowledge about effective methods and techniques of summarizing information embedded in digital images, but also has numerous practical applications in photo management systems, image search engines, photo library optimization processes, and smart cameras. Previous work has explored the method of using generic local descriptors extracted from gray-scale photographs to assess the aesthetic quality. In our work, we designed an aesthetic descriptor based on color information of local patches to asses the aesthetic quality of photographs. We employed the Bag of Words (BOW) model to describe our data. The BOW model is widely used in text analysis tasks due to its good performance. This model can also be adopted by computer vision tasks to classify and categorize images. With such a model, every photograph in our dataset was represented as a vector of visual words derived from the color descriptors. A Gaussian mixture model was used to cluster and verbalize photographs, which is proven to be highly effective in our classification task. The best performance we achieved in our experiments was 72.9% AUC.