Research On Quality Evaluation Of Unreferenced Images Based On Deep Learning

Non-reference image quality assessment(NRIQA)also known as blind image quality assessment(BIQA)is means that the quality of the image is directly evaluated without a reference image or without the reference image features as a reference.Its purpose is to simulate the visual system of the human eyes to achieve image quality assessment without reference images automatically and accurately.It is an important research content in the field of computer vision and image processing,and is a key technology for automatically judging and capturing high-quality images.Since deep learning has achieved significant success in the field of computer vision,it has become a major contributor to this field.It can learn complex features that correctly represent targets from a large number of samples based on a data-driven approach.Although the theory based on statistical principles needs to be improved,the practice has continuously made new breakthroughs and is rapidly used in the commercial field.A large number of network models have also emerged for image quality assessment problem.In order to achieve the image quality assessment without the reference image or the reference image features in accordance,these models learn quality features from rich distortion samples,use typical models such as VGGNet,residual networks to get image quality features by transfer learning,or combine semantic information to describe the features of image quality.Based on the deep learning method,this paper starts from the actual demand of no reference image quality assessment,and focuses on the fast and accurate real-time image quality assessment without reference,and conducts in-depth research based on the convolutional neural network optimization strategies.The main research contents of this paper are as follows:(1)Research on non-reference image quality assessment based on shallow convolutional neural network: In order to obtain a simple and accurate non-reference image quality assessment model,a comparative analysis of typical methods for nonreference image quality evaluation based on deep learning methods from 2014 to the present is carried out.The mainstream quality assessment structures which are divided into single task structure and multi-task structure,and the multitasking structure contains three types of structures.Then,based on the shallow convolution network,the single task model and the multitasking model are implemented and compared experimentally.In additions,through the multi-task model sub-task output dimension experiment,it can explain that in the non-reference image quality assessment research,it can pre-train on the relevant data set according to the sub-task requirements and objectives,and combine the quality assessment task fine-tuning,with the transferable learning integrated into other tasks.Advantages.(2))Research on deep non-reference image quality assessment based on distortion level: In order to extend the non-reference image quality evaluation model to the actual application scenario which can be applied to both single-distortion images and multidistortion images,a non-reference image quality assessment algorithm based on distortion type is proposed.It is based on the study of shallow neural network models for multitasking structures.At the same time,in order to explore the expression of image quality features of more and deeper networks,the network depth is extended.The extended model is first pre-trained based on the distortion level,and the quality prediction training without reference images is performed.In additions,this paper uses SROCC and PLCC two performance metrics to measure the accuracy and the correlation with the human eyes,and compared with the typical methods IAQ-CNN++,DIVINE,BIQI,etc.The experimental show that the accuracy of this method is 0.924,and the correlation with human eyes is 0.931,at the same time,the parameter size is much smaller than other models.Finally,based on the research results,the optimization strategy for the nonreference image quality evaluation problem is proposed from different directions.