Image Quality Assessment Based On Few-shot Learning

With the rapid development of digital media technology and mobile smart devices,the amount of data in digital images and videos is growing exponentially,but digital images may be distorted at various life cycle stages such as acquisition,compression,storage and transmission,which not only affects the user’s sensory experience,but also may affect the usefulness of the information obtained by the user.Therefore,in order to ensure the user’s sensory experience and the usefulness of the information obtained,it is very important to find an accurate and efficient method for evaluating image quality.At present,the mainstream image quality assessment method is mainly based on deep learning,and deep learning often relies on massive data for training to achieve the desired effect,and there is a problem of insufficient manual labeling samples in the field of image quality assessment,resulting in deep learning model overfitting and lack of certain generalization capabilities.Few-shot learning can enable network models to learn richer visual information expressions using a small number of labeled samples,and solve the problem of insufficient samples in the field of image quality assessment.In this paper,the image quality assessment method based on few-shot learning is studied,and its main research content and innovation are as follows:(1)A referenceless image quality assessment based on few-shot transfer learning is proposed.First,a new dataset is constructed by a self-supervising task and used to train a feature extractor that can predict the rotation angle of the image;then the feature extractor is transferred to the new task by using few-shot transfer learning;and finally,a local perception module for the problem of local distortion of images is added.This method solves the problem of insufficient manual labeling samples in the field of image quality assessment,and considering the local distortion of images.Experimental results show that the proposed method achieves good results on different image quality evaluation datasets.The SROCC value in the LIVE dataset reached 0.973,surpassing most methods.(2)A referenceless image quality assessment based on few-shot semi-supervised learning is proposed.Firstly,a large amount of soft label data is generated by using the full reference image quality assessment method and the SVR model through the self-supervised learning task;then the same feature extraction network is jointly trained by two fully connected channels;finally,the splicing feature extraction network and the fully connected layer trained with real label data are used as the final image quality assessment network to evaluate the images.Experimental results show that this method has achieved good results and has good generalization ability.The SROCC value reached0.871 in the totallenge dataset,showing the best results of all the methods in this paper.(3)A referenceless image quality assessment algorithm based on few-shot meta-learning is proposed.Firstly,a feature extractor with prior knowledge is trained by few-shot characterization learning after generating soft label image data by combining the self-supervised learning task;then the network model is trained by using the meta-learning method,and the learning strategy combined with a difficult task reprocessing in it forces the network model to continuously adjust closer to the global optimal parameters;and finally,it is fine-tuned on the training set of the target dataset and used as an image quality assessment task.Experimental results show that the method in this chapter can effectively predict the mass fraction of images,and the good results on different data sets show that the method has good generalization ability.