| The popularity of digital camera and social network has greatly enriched our spiritual life, and we can easily access vast amounts of digital photos. However, many photos taken by amateur photographers are not beautiful for the lack of professional instruction and bias in aesthetical appreciation. This paper focuses on image aesthetics, we use computer vision approach to simulate human perception, and propose some methods to assess and to enhance image aesthetics. It can be widely used in many fields, such as image aesthetic assessment, photographic auxiliary amendment and image aesthetic enhancement. There are two major tasks in image aesthetics, i.e. image aesthetic assessment and image aesthetic enhancement. The procedures include extracting features from images that represent the inherent image aesthetics, and designing feasible measurement related to aesthetic theories. These features are finally used to evaluate and enhance image aesthetics. Our main achievements are followed: 1. In this paper, we design a series of image aesthetic features, which include image visual features and image semantic features. Image visual features are designed according to basic photography rules, which include global and regional features. Moreover, image visual features require much priori knowledge and engineering practice, and these visual features cannot contain all photograph techniques. Therefore, in order to further improve the performance of the aesthetic evaluation, this research applies the LLC coding, which is an efficient method in the field of image classification and can identify the semantic features effectively. The experimental results show that our semantic features can be used as auxiliary information, which can improve the efficiency of image aesthetic assessment. 2. In this paper, we purpose a new content based image aesthetic enhancement. Since different image aesthetic distortions exist in different types of images, we design four enhancement strategies for four image categories, i.e. human, architecture, night images and other images. Four enhancement strategies are followed: 1) for the image of human face, we detect face regions in the image firstly, and then use the bilateral filter to “buffing†the face regions in the image to smooth face regions. 2) For the image of architecture, we detect long lines in the image. Based on geometric transform, we use affine transform to rotate image and keep the architecture stand straight on the ground. 3) For the image of night, we introduce a deblur method to remove motion blur caused by arm shake during a long exposure time. 4) For the rest kinds of image, this paper proposes an optimal method to enhance the quality of image composition. The experimental results have proved that the four enhancement strategies are greatly efficient. |
