Research On No-reference Image Quality Assessment Based On Deep Learning

No-reference image quality assessment can directly evaluate the quality of distorted images objectively without reference images,and has important practical value for image optimization algorithms and image quality control.This article takes distorted images without reference images as the research object,and follows two approaches: improving the model and changing the training strategy.We respectively study how to improve the accuracy of image quality prediction by improving the distortion feature extraction ability of deep learning models,and how to use deep meta learning training strategies to improve model prediction performance under unfavorable conditions in small datasets and natural distorted datasets.Its main tasks are as follows:(1)In order to further improve the accuracy of no-reference image quality assessment,this dissertation proposes a no-reference image quality assessment model based on dual branch convolution and attention.Firstly,an edge extraction operator is used to extract edge images from the original distorted images.A specialized network branch is designed to extract edge features from the edge images,and the main branch network based on Inception Res Net is improved by introducing an attention module to extract more effective and rich distorted features.Subsequently,the two features are fused,And use fully connected layers to establish a mapping relationship between convolutional layer features and the image quality score that needs to be predicted.The experimental results show that the PLCC index and SRCC index of this method on the image quality evaluation dataset Koniq-10 k are 0.945 and 0.928,respectively,improving the accuracy of image quality evaluation.In response to the problem of the aforementioned heavyweight models occupying a large amount of storage space,which limits the deployment and application of low-performing devices,using model quantization technology can save about three times the storage space when the recognition accuracy decreases by less than 2%.(2)The dissertation propose a No-reference image quality evaluation method based on hidden gradient model independent meta learning algorithm(model-agnostic Meta-Learning with implicit,i MAML).The dissertation utilizes the special task construction in meta learning training to associate it with distortion types,allowing the meta learning algorithm to learn on specific distortion type tasks,expecting the model parameters to adapt to various distortion types.Through the internal and external two-layer optimization process of meta learning algorithms,each distortion type is treated as an optimization task.During the optimization process,not only the optimal parameters of a single task are learned,but also the average optimal parameters of multiple tasks are learned,obtaining good parameter optimization directions.In the face of unknown distortion type tasks,the model can quickly converge,reducing the dependence on training samples.And when faced with unknown distortion types and natural distortion images,the prediction effect is good,and pyramid pooling technology is also used to solve the problem of accuracy degradation caused by excessive image scaling.The experimental results show that the algorithm achieves higher recognition accuracy than mainstream methods on multiple datasets.