Adaptive Optics Image Quality Assessment Based On Deep Learning

The adaptive optics imaging system is affected by atmospheric turbulence,vibration,and the system’s photoelectric noise during the imaging process,resulting in blurred imaging results.Although the adaptive optics system can correct most of the wavefront aberrations,the adaptive optics imaging system still leads to uneven imaging results due to residual atmospheric turbulence,closed-loop tracking errors,and optoelectronic noise.Poor quality of adaptive optics images is not beneficial for postprocessing of data and post-mining of data,so the quality of adaptive optics images needs to be assessed.At present,experienced observers are required to subjectively screen out good quality adaptive optics images,but the process of subjective screening is labor-intensive,and the process of subjective screening can be affected by factors such as observer emotion.Secondly,there are large differences between adaptive optics images and natural images,which leads to traditional image quality assessment methods being unreliable for adaptive optics images,and even the assessment results may deviate from the actual situation,so it is necessary to study objective quality assessment methods for adaptive optics images.The main research work and contributions of this dissertation are as follows.Theoretically analyzes the objective image quality assessment methods currently mainly used to assess the quality of natural images,and finds that the features extracted by the quality assessment method based on deep learning are more diverse and can fully extract useful information.Currently,the image quality assessment method based on deep learning has achieved better accuracy than the traditional method.Based on this,this paper proposes a quality assessment method for adaptive optical images based on deep learning,and applies deep learning to assess the quality of adaptive optical images.However,there are two difficulties in the implementation of the method proposed in this dissertation: First,the training of deep neural networks requires a large number of adaptive optics images;second,generating image quality labels for a dataset containing a large number of adaptive optics images.In order to obtain a large number of adaptive optics images required for deep learning,this dissertation first reconstructs a 3D model of the observed object from known parametric data of the observed object.Project the 3D model onto the 2D plane to form a reference image.In the process of projection,reference images with different target poses and different contrast are generated by changing the projection angle and light orientation.Secondly,by simulating the imaging process of the adaptive optics imaging system,reference images are input into the adaptive optics image degradation model to obtain simulated adaptive optics images with varying degrees of blur,and finally,the adaptive optics image dataset containing 400,000 frames is formed.This dataset will be used for the subsequent training of the deep neural network.To solve the problems of high workload,subjectivity and unstable assessment results of the subjective way to generate image quality labels,the dissertation uses a reference image quality assessment method and other features that can be used to assess the quality of adaptive optical images,such as contrast and target occupancy,to jointly generate the quality labels of images,and finally generates quality labels that match the subjective judgment of the human eye.As a result,a dataset with quality labels for assessing adaptive optics image quality is obtained in this dissertation.A neural network model is built with a classical network,such as the Res Net series,as the backbone of the network.The generated adaptive optics image data are fed into the neural network for training and the neural network model for assessing the quality of adaptive optics images is obtained.The model has a Spearman correlation coefficient(SROCC)best of 0.994 and a Pearson correlation coefficient(PLCC)best of 0.993 on the test set.The experimental results show that the method in this dissertation considers a variety of degradation factors in the process of adaptive optics image imaging in a comprehensive manner.A no-reference adaptive optics image quality assessment model is obtained by training a deep neural network.The assessment accuracy is better than other traditional image quality assessment algorithms.And the model is also feasible for assessing real acquired adaptive optics images,proving that the model has good generalization performance.