Research On Computational Image Aesthetic Evaluation And Optimizing Composition

Computational image aesthetics aims to simulate visual perception and aesthetic thinkingof human being, so as to perform aesthetic evaluation and enhancement in accordance withhuman judgment. Computational image aesthetics is an interdiscilpinary subject thatpertinents to art, psychology, computer science. It can be widely used in many fields includingimage quality assessment, computer aided evaluation and enhancement of photo, automaticanalysis of painting style and so forth. Computational image aesthetics has two major tasks，including how to extract features from images that represent the inherent image aestheticvalue, and design feasible measures related to aesthetic theories. According to these featuresand measures, we can use suitable algorithms to obtain the results of aesthetic evaluation orenhancement. In this paper, we focus on image aesthetic evaluation and optimizingcomposition and we have following achievements:1. We propose an innovative image saliency detection algorithm. First, we develop aself-adaptive approach for image pre-segmentation. It determines the number ofsegments according to the complexity of an image so as to produce homogeneoussegments. Besides，texture and color features are extracted and we incorporate themusing a nonlinear method. The combined feature can depict the difference betweensegments more accurately so as to highlight the prominent segments. Moreover, with thecombination of distinctive and spatial descriptors, it is more accurate to represent therelation between each segment. Finally, we adopt a Harris point detector to obtain priorknowledge of salient regions, which strengthens the saliency of subject area andsuppresses the background. It help us to produce a more precise and clear saliency map.2..We design a comprehensive image aesthetic evaluation model. It comprises two parts:classification, and score prediction. First, we extract features including low-level visualfeatures, high-level features and regional features. Some of them are new developed onthe basis of previous work, or newly introduced according to aesthetic theory. Undermachine learning scheme, we build the aesthetic classification model and score precidtion model by using an SVM based classifer and an SVR algorithm. The promisingresults demonstrate the efficacy of our model on aesthetic evaluation.3. We present an optimizing composition algorithm based on image retargeting techniques.This algorithm defines several energy mesh warping techniques to build the frameworkof image composition. A set of aethtetic energy and shape preserving equations areproposed to preserve the salient objects and move to optimal. We use a quadraticprograming scheme to obtain the optimal position for each control point. An imagedeformation method is used to obtain the final results. Our experiments show that ourmethod is applicable and effective for image aesthetic enhancement.In this paper, we mainly focus on the tasks of image aesthetic evaluation and optimizingcomposition. In addition, we make contribution to image saliency detection, which is a crucialpart for above tasks. To deal with some existing challenges in these studies, we propose somenovel methods, and we make some impressive achievements.