Research On Deep-feature Based Image Quality Assessment Algorithm

The proliferation of social networks and digital cameras has led to an explosive increase in the number of digital images being created and distributed.This has resulted in large datasets of photographs for creating and maintaining personal memories of events and places.However,construction of an ideal personal album or collection manually from such massive number of images is onerous,and the task can often be time consuming and challenging.In order to improve the low accuracy and unawareness to the user's preference of the existing method,with challenge described above considered,the paper proposed a deep-feature-based image quality assessment algorithm.The proposed algorithm contains two part:1)Personalized ranking model.2)User-specific aesthetic attribute distribution model.In module 1,our ranking algorithm first takes as input a series of photos that users prefer.Then the powerful DCNN is deployed to extract the deep feature from user-specific images and compared with that from training datasets to generate user-specific training dataset.Finally,SVMrank is applied to learn a customized hyper-plane and then a personalized ranking is released.In algorithm 2,we first extracts deep feature and applies feature encoding strategy via using powerful RBM.Then,combined with several aesthetic attributes,a multi-attribute classifier is trained.Finally,our system takes as input a series of user-specific images and produces as output a reliable,user-specific aesthetic attribute distribution.In the proposed personalized ranking model,the learned ranking function is deduced from SVM^rank.After acquiring the score of the ranking,our system deploys normalization and weighed verify over the score above and decision value to generate a final ranking.In the proposed user-specific aesthetic distribution model,the system focuses on identifying which aesthetic attributes the user is interested.To that end,the paper scratches several kinds of images with total amount of 60005images attached with label and score in professional photography website.Then the images is divided into several aesthetic kinds according to the professional photographic rule such as horizontal rule of thirds and rule of thirds to learn a local feature?The chosen aesthetic rules?and global feature?Color?classifier.Finally,a user-specific aesthetic distribution model is learned by combining deep aesthetic feature and labels.During testing stage,the paper conducts extensive experiments and user studies on two large-scale public datasets?CUHKPQ and AVA?and compares the results of proposed system with that of related work.Besides the use of conventional photographic images,the paper also uses UAV aerial shoots as input,and the relevant experiments are performed to verify the effectiveness of algorithms.Experimental results demonstrate the effectiveness of the proposed algorithm.For personalized ranking part,the model enables an effective ranking that meets user's preference.For user-specific aesthetic attribute distribution part,the model produces a reliable aesthetic distribution after acquiring user's choice.Based on this,combining with personalized ranking algorithm and personalized aesthetic distribution algorithm,the proposed user-specific aesthetic assessment could provide moderate reference for exploring the personalized preference.