Research On Optimal Series Photo Selection Method Based On Multi-scale Visual Attention Mechanism

With the popularity of smartphones and cameras,people are gradually getting used to recording every unforgettable moment through photos.To ensure the best shot is captured,users usually take a series of photos of the same object or scene.However,due to space saving considerations,the user must manually determine the retained objects from these almost identical images.This process is often time-consuming and laborious.More and more researchers hope that computers can assist users to select more attractive photos from similar photos,that is,to find an automatic evaluation method that promotes the selection of optimal images in a series of photos.Most of the previous research work on image aesthetic quality evaluation directly used the features generated by the top convolutional layer to classify or regression predict the image These methods only use the features of neural network with high-level semantic information.Processing similar series of photos tends to produce similar predictions.As a result,the model performance is poor.This article is dedicated to transforming the aesthetic quality evaluation of a series of photos from a simple classification or regression task to an image sorting and comparison task.The objective of aesthetic evaluation is achieved by sorting the images on the aesthetic levelBased on the above background,in the work of this thesis,an optimal series of photo selection methods based on the Multi-Scale Visual Attention Mechanism(Multi-Scale Visual Attention Mechanism)is proposed.First,this method uses a twin-convolutional neural network structure with shared parameters to learn the distinguishing characteristics of two similar images to make a relative aesthetic quality assessment of the image pairs.Secondly,we add lower-level features to supplement high-resolution image details that high-level semantic features do not have,and analyze the effects of different convolutional layer output features of the convolutional neural network on aesthetic quality evaluation through experiments,and finally aggregate The output features of the last three layers of the network replace the top-level features of previous studies to predict image beauty.Finally,because the use of lower-level features will bring a lot of redundant information,this thesis also uses the parallel attention mechanism and the self-attention mechanism to filter out the interference information to obtain a more streamlined aesthetic feature,which has a good generalization and has an integrated global and Local information capability.The proposed method can accurately and adaptively extract the characteristics of high-resolution information and high-level semantic information from a series of photos,which provides a new idea for processing the aesthetic evaluation of series of photos and traditional images.Moreover,this method is the first to introduce multi-scale feature fusion technology in the field of image aesthetic quality evaluation to obtain high-resolution detail features,and uses attention mechanisms to filter the spatial and channel dimensions of features.This article shows in detail the design of the network framework and the role of various technologies in the structureIn order to verify the effectiveness of the method,this thesis conducted a more systematic experiment on the Atrium dataset containing 5953 series of aesthetic datasets Phototriage and the improved AVA dataset containing 7083 images.The method proposed in the thesis has an accuracy rate of 76.47%and 93.13%respectively in the above two data sets,and its performance is better than the latest research methods of image aesthetic quality evaluation as a whole.In addition,we use the ablation experiment to analyze the necessity of network composition and structure,and verify the rationality of the structure design of this thesis.