Image Aesthetic Assessment Method Based On Multi-Feature Fusion

Aesthetics covers psychology,linguistics,anthropology and other fields.It is a subject that studies the relationship between human and aesthetics.The traditional way of aesthetics is given by the subjective opinion,but this approach requires people to have profound aesthetic attainments and related fields of knowledge.With the development of computer vision,image processing and pattern recognition technology,and the explosion of the number of digital images,the standard of image aesthetics has been gradually quantified.In this context,the concept of computational aesthetics in the field of computer graphics arises at the historic moment.Image aesthetics can be calculated by constructing a computational model,simulating the human brain's aesthetic thinking and visual system,to make a feasible aesthetic decision.Aiming at the problem that the feature selection of the existing methods is single and can not be applied to all types of images,this paper proposes an image aesthetic evaluation method based on improved color harmony and multi-feature fusion.The existing color harmony calculation method uses a fixed block size,and ignores the differences in the subject area and the background area.The method proposed in this paper first detects the salient region of the image,and then separates the subject from the background,and divides the two regions according to the entropy of color information.On this basis,SLIC algorithm is used to determine the main color in the block.Finally,the color harmony feature is calculated according to the Moon-Spencer color harmony theory.At the same time,by the color harmony feature,color,composition,sharpness,texture and the statistical properties of discrete cosine transform of the image are merged to constitute feature vectors of the computable aesthetic evaluation model and trained by SVM classifier to get the result of aesthetic evaluation.In the experiment,the method of this paper is tested by using Datta database,PhotoQualityDataset database and self-built YUAestheticData database,and the classification of the image's beauty is realized.In the Datta database,the average classification accuracy rate was 82.04%;in the PhotoQualityDataset database,the average accuracy rate was 84.8%;in the YUAestheticData database,the average accuracy rate was 83.91%.By setting different experiments and comparative experiments,it is shown that the computable image aesthetic evaluation method proposed in this paper has good accuracy by improved color harmony feature calculation,and can be applied to many types of images at the same time.