Image Aesthetic Quality Assessment Based On Deep Learning

The assessment of image aesthetic quality is to automatically make an objective evaluation of the aesthetic perception of image by simulating human perception and understanding of beauty using computer.This problem is usually divided into two categories:automatically classifying an image into high or low aesthetic quality and the prediction of image aesthetic quality score.Image aesthetic assessment is a challenge task in the computer vision community because of the complexity and subjectivity of human's aesthetic activities.In recent years,the development and application of deep learning has achieved unprecedented results,specially deep convolutional neural network(DCNN)has been proved to be effective in many computer vision problems.Therefore,in this paper we adopt DCNN that conduct the image aesthetic quality assessment.We propose a novel DCNN structure codenamed ILGNet that connects intermediate Local layers to the Global layer for the output.In this model,we can extract the image local features in front of the network layers and extract the image global features at the back of the network layers.The model evaluates the aesthetic quality of images using the two kinds of features extracted.In addition,we extends to multi value regression problem about the unit regression assessment of image aesthetic quality.This paper proposes a new method of score distribution assessment and designs a new loss function(RS-CJS).This loss function can be well applied to the multiple valued regression problem.At last,we train and test this model in a large-scale dataset for aesthetic visual analysis(AVA).The experimental results show that the proposed ILGNet outperforms the state of the art results on the classification and regression problems of the assessment of image aesthetic quality in the AVA benchmark.