Research On Aesthetic Models Based On NAS Neural Architecture Search

With the advent of the 5G era,the data flow and network speed are no longer bottlenecks.Human beings are constantly exposed to a large number of pictures and videos through computers,mobile phones and other platforms.There are massive amounts of pictures and videos every day in e-commerce platform and other applications like traveling and video social.How to quickly and accurately recommend beautiful and attractive pictures or video cover pictures in this huge amount of data will greatly determine the corresponding data flow and revenue.Screening was performed manually in the past,but in the face of large amounts of data,relying on manual labor has become inadequate.How to let the computers to replace manual labor to complete complicated picture screening has become the first choice and it is becoming a hot topic nowadays.Computable aesthetics,that is,the habit of learning human aesthetics through computers,eventually replacing humans to score pictures.Most of the related research at the beginning was to simulate human aesthetic habits,by capturing artificially designed features such as lighting,brightness,and color as the scoring basis,and using a classifier such as SVM to score.With the development of deep learning,researchers have found that artificially designed features are not comprehensive,and deep convolutional networks can capture a lot of feature laws that humans have not discovered.However,the current computing power of mobile terminals is not enough to support large convolutional networks,so convolutional neural networks are also developing,and new artificially designed small parameter networks have appeared.With the rise of NAS architecture search,letting machines replace artificial experts to automatically find suitable neural network structures has become a new upsurge.This thesis combines architecture search technology and computable aesthetics technology for the first time.NAS architecture search technology is introduced into the field of aesthetic models.Through a progressive and differentiable search strategy,a new simple and efficient search is performed on the AVA aesthetic data set and discovery the aesthetic model named Aesthetic Net.In addition,most of the current aesthetic models are based on predicting a final output that is evenly divided.Finally,the image is classified into good or bad by setting a threshold.The approach in this article is different as it is to simulate the multi-person score by outputting the distribution of the score.Based on the obtained Aesthetic Net,this paper combines aesthetics-related theoretical analysis and neural network vision-related technologies,and proposes some effective improvement measures.In order to solve the problem of large amount of convolutional network parameters,this paper uses a deep separable convolution module in the search.After observing the difference between the distribution of the predicted results and the actual distribution,this paper proposes a self-weighting mechanism to calculate the loss function.To further simulate the focusing process of the human eye,this paper introduces a two-dimensional attention mechanism to increase the recognition of the network.Considering the impact of the integrity of the picture on visual aesthetics and in order to overcome the drawbacks of the convolutional neural network fixed input size,we use adaptive pooling to ensure the original proportion of the input picture,and use hole convolution to expand the receptive field of the network,so that we can obtain more overall information when learning.Finally,this article designed multiple sets of comparative experiments to verify the effectiveness of the aesthetic model Aesthetic Net obtained through the NAS architecture search technology and the effectiveness of the proposed improvement strategies.