Image Aesthetic Evaluation Based On Latent Semantic Feature

The image aesthetic evaluation aims to make the computer assess the aesthetic value of images as the simulation of human subjective aesthetic thinking,which is a very difficult challenge in the field of computer vision and image processing,there is no unified evaluation standard.Due to the simple features,there exists the problem of low correlation with human subjective perception.Most of image aesthetic evaluation methods only pay attention to the hand-crafting feature extraction,and the evaluation effect depends heavily on the performance of object region extraction which leads to low aesthetic classification accuracy.This thesis proposes a computational image aesthetic evaluation model based on latent semantic feature.Firstly,the local hand-crafting features are extracted on the superpixel block image weighted by point-ofinterest density,secondly the latent semantic feature and semantic feature are obtained by mapping and encoding hand-crafting features.Finally,refer to the image aesthetic classification evaluation model established by machine learning method,the image aesthetics value can be obtained automatically.The main work and novelty of this thesis are followed:Firstly this thesis proposes a new image aesthetic evaluation feature--latent semantic feature.Similar to the latent semantic analysis in the field of documents,the latent semantic feature is a new aesthetic feature that measures the similarity and co-occurrence of aesthetic feature vectors.Through encoding the aesthetic feature vector by word mapping,we can obtains the aesthetic features document.By calculating the topic words of the aesthetic feature document using the Latent Dirichlet Allocation(LDA)model and calculating the similarity of the aesthetic feature document vocabulary and the topic words,the latent semantic feature is obtained.Second this thesis designs a super-pixel segmentation algorithm based on the density of interest points to extract the image local hand-crafting features.It replaces the SPM(Spatial Pyramid Matching)algorithm to introduce image space attributes,greatly reduces the feature dimension.The extracted image features are related to the aesthetic complexity attributes through weighting by the density of interest points,which measures the complexity of the local area of image.Lastly in order to further improve the effect of aesthetic classification model,this thesis uses Locality-constrained Linear Coding(LLC)method to extract the semantic features of aesthetic images.The final image aesthetic evaluation model is obtained by combining latent semantic feature and semantic feature.The experiments are performed on the AVA database,using the accuracy confusion matrix and ROC curve as evaluation indicators to measure the performance of the classifier.Comparing with other aesthetic evaluation methods,the results show that the proposed latent semantic feature is effective for the image aesthetics evaluation,and the combination of latent semantic features and semantic features can improve the accuracy of the image aesthetic classification model.