Recently, it is expected that the systems for automatically adding the suitable music to a target image will be developed. Toward realizing the systems, we propose a method to estimate the degree of feeling that natural objects appear in the target (photographic) images. The method uses neural network-based frameworks, and the network is trained by approximately 900 images. Evaluations were conducted by a task which classifies the impression of input images as “natural” or “artificial”. The best classification rate of 76.5% is observed when using “shape information (Bag-of-Keypoints)”, “the number of object candidates”, and “the number and area of detected face regions” as input features.
View full abstract