Research On Image Aesthetic Quality Evaluation Based On Multi-feature Fusion

With the rapid development of the Internet and the popularization of camera devices such as mobile phones,the number of images has grown exponentially,and the demand for image aesthetic evaluation by users has increased.Image aesthetic quality evaluation aims to simulate human vision and aesthetic thinking,and then perform aesthetic modeling of images,so that computers can automatically evaluate the aesthetic value of images quantitatively.The core is to establish the relationship between image content expression and image aesthetic determination.Image aesthetic quality evaluation research has important application value in face recognition,target detection and image search.It has attracted a wide range of attention from researchers in recent years.In general,most of the current research methods for image aesthetic quality evaluation use deep convolutional network models to automatically extract features for aesthetic quality evaluation.However,users usually evaluate the aesthetic quality of an image based on understanding the image semantics,that is,when evaluating an image aesthetically,users would first perceive the image content and then evaluate it.Therefore,the image aesthetic quality evaluation should be linked to the image semantics.However,most of the existing researches carry out the two tasks of image aesthetic quality evaluation and semantic recognition separately;on the other hand,the aesthetic quality evaluation of images is subjective in nature,and the aesthetic scores of images in the existing aesthetic quality evaluation datasets are generally determined by different people's scores.Therefore,it is also very important to simulate the subjective evaluation process of people in the design of the image aesthetic quality evaluation network model.From the above two perspectives,this thesis proposes two image aesthetic quality evaluation methods based on multi-feature fusion.The main contributions and innovations of this paper are as follows:(1)An image aesthetic quality evaluation method fused with semantic features is proposed.Image aesthetic quality is closely related to image semantics.Therefore,the two tasks of image aesthetic quality assessment and semantic content recognition are interrelated and share certain features,and the performance of image aesthetic quality evaluation can be effectively improved by associating the two tasks of image aesthetic classification and semantic recognition as a whole.This method uses a multi-task parameter soft sharing method to formulate learning strategies,and combines aesthetic quality evaluation and semantic content recognition,then effectively learns the relationship between tasks and the characteristics of each task.Specifically,this method learns global information in the channel dimension and local information in the spatial dimension,and then performs feature adaptive fusion between tasks in the two dimensions so that the effective features of each task can be enhanced and the invalid features can be suppressed.Finally,the stability of the task learning process is ensured by adding self-learning parameters so that the performance of the two tasks can finally promote each other to improve.(2)An image aesthetic quality evaluation method based on multi-feature fusion is proposed.The evaluation of image aesthetic quality is closely related to human subjective cognition.Therefore,this method uses multiple deep convolutional neural network classifiers to learn different representations and preferences of image aesthetics by simulating the process of evaluating the aesthetic quality of the same image by different users,thus forming an image aesthetic evaluation model by multi-branch feature fusion.In the process of learning and training,the model of this method interacts with the features learned by different aesthetic feature extractors through the multi-branch feature fusion module,that is,a feature fusion mechanism is added between multiple branches deep convolutional neural networks,which can generate more detailed and comprehensive aesthetic features,more effectively establish a close connection between image content expression and aesthetic judgment,and ultimately achieve an effective improvement in the performance of image aesthetics quality evaluation task.