Multi-view And Adaptive Image Aesthetic Quality Assessment

Image aesthetic quality assessment has been a hot topic in the computer vision field. This research tries to automatically judge whether a given image has "high" or "low" quality from the view of human aesthetics. The image aesthetic quality assessment method can be applied in other computer vision systems, such as image retrieval, image management and human-computer interaction system, to improve the user experience. In this paper, we conduct careful research on image aesthetic quality assessment and make four contributions:1). Design a new group of rule-based aesthetic features; 2). Apply the deep learning features on the image aesthetic quality assessment; 3). Merge aesthetic features which come from multiple views to improve the performance of image aesthetic quality assessment; 4). Propose the adaptive quality assessment model to improve the framework of the current image quality assessment research.1). Design a new group of image aesthetic features, which are more effective and have lower computational complexity, based on the photographic knowledge:Through careful analysis on the image attributes that are related with image aesthetic quality, we propose a new group of features. Compared with current aesthetic features, these new features perform much better;2). Apply the deep convolution neural network on image aesthetic quality assessment:Deep convolution neural network is proposed by Krizhevsky et al. and Hinton et al. to handle with the generic image classification problem. This model has similar structure with the human vision system. It takes the image as input and can automatically learn the image content. Due to its strong learning power, this network has been used in many other computer vision problems and has achieved amazing performances. Here we apply it on the image aesthetic quality assessment and also get outstanding results.3). Use the multi-kernel learning method to merge different aesthetic features:The deep learning feature compresses much image content information inside. It indirectly describes the image aesthetics and is complementary with the rule-based hand-crafted feature which directly describes image quality. Here, we adopt MKL method to merge these two kinds of features to better handle with the image aesthetic quality assessment problem. This is also the first time that people try to merge aesthetic features that come from different views;4). Improve the current aesthetic quality framework use the characteristic of the image aesthetic quality assessment problem:According to photographic knowledge, different camera settings and photographic skills will be adopted on different subjects. Therefore, the aesthetic quality metric should be different for different kinds of images. Here, we propose the query-dependent image aesthetic quality assessment model. Whenever given a test image, we first use content-based image retrieval methods to get a group of similar images from the training dataset. Then, an assessment model will be specifically trained for the test image which is equal to find the proper aesthetic quality metric for it. The effectiveness of this model is verified through many experiments.This paper makes many contributions on the image aesthetic quality assessment and conducts many experiments to verify these thoughts. Wish this paper can be helpful for other researchers.