Research on Image Aesthetic Quality Evaluation Method Based on Deep Feature Fusion
The evaluation of image aesthetic quality aims to use computer simulations to simulate human perception and cognition of beauty,and automatically evaluate the"beauty"of images.Although existing deep learning based aesthetic quality evaluation methods have achieved good performance,they typically use advanced features from convolutional neural networks(CNNs)for aesthetic prediction,while low-level and mid-level features highly correlated with aesthetic prediction are still overlooked.This article proposes an aesthetic quality evaluation method based on deep feature fusion,which extracts features from different layers of convolutional neural networks for feature fusion,and uses the fused features to predict the aesthetic score of images.In order to evaluate the effectiveness of deep feature fusion algorithms,this paper constructs three deep feature fusion networks based on MobileNetV2,VGG19,and Inception-v3 networks,and conducts experiments on the AVA dataset.The experimental results show that compared to the high-level features in the original network,fused features can significantly improve aesthetic prediction performance.
NETL
NSTL
万方数据
