Research On Image Aesthetic Evaluation Using Deep Convolutional Neural Networks

With the creation and acquisition of digital images becoming more and more convenient, the number of digital images shows explosive growth. Every day, countless images are shared through Internet. The image management becomes time-consuming and burdensome due to the sharp increasement of image amount. People tend to capture and store the high quality images. Many areas or tasks, such as image retrieval, image design, style analysis of artistic works, human-computer interaction tasks and so on, need aesthetic assessment of images. Thus, automatic evaluation of image aesthetics has become a hot topic in the field of image processing.This paper mainly focuses on how to evaluate image aesthetics automatically, which is using computers to judge the aesthetics of images in the perspective of human thought. This paper is to classify images into high aesthetic and low aesthetic categories. The traditional image aesthetic evaluation methods are as following.(1) First, design and extract handcrafted features aimed at image aesthetics, then use the extracted features for classification via machine learning algorithm.(2) Extract the widely known generic features of images, then use the machine learning method for classification. Both methods can't expound aesthetic information fully and accurately.In recent years, the study of deep learning has attracted enormous attention and acquired breakthroughs in many computer vision tasks. Deep learning method analyze image matrixes directly, it can learn features through deep neural networks automatically. This paper attempts to assess image aesthetics using deep learning, mainly focus on image aesthetic classification via deep convolutional neural networks. And combine handcrafted features with features extracted from convolutional neural networks. The methods in this paper have got better classification results than existing image aesthetic classification methods.In summary, this paper has the following contributions.1. Aimed at image aesthetic problems, this paper designs paralleled deep convolutional neural networks. Understanding the features learning ability of deep learning networks deeply, and considering the impact of color, brightness, composition and other image information, this paper designs different image description matrixes as the inputs of paralleled networks. Then we combine these features effectively, in order to obtain image features that can express more exhaustive aesthetic information. From the experimental results and their comparison, the parallel convolutional neural networks algorithm obtains better image aesthetic classification results than existing handcrafted features, generic features, and the latest deep learning methods.2. Handcrafted features are often designed directly from the image characteristics, aesthetic rules, visual psychology and so on. This paper combines handcrafted features and features of convolutional neural networks, in order to express image aesthetic information fully from different aspects. In the image aesthetic datasets, train the combinational features with SVM(support vector machine) method to obtain the image aesthetic classifier. Compared with the latest image classification methods, the combination of these two types of features performs better.