Research On HDR Image Quality Assessment Based On CNN

High Dynamic Range(HDR)images have strong performance capabilities in dynamic range of scene and have been widely used in aviation remote sensing and medical imaging.However,due to the limitations of acquisition,transmission and display devices,qualities of HDR images at the receiving end are degraded.So,it has important significance to establish an effective HDR image quality assessment(IQA)method.At present,traditional machine learning methods such as support vector machines and k-means can be used in HDR image quality assessment,but traditional machine learning techniques only use shallow frameworks,and it is unable to highly imitate the mechanism of human visual perception to judge image quality.Hence its application is limited in image quality assessment area.The convolutional neural network(CNN)has been widely used in HDR image quality assessment field because it can better capture basic features of the image.Based on the in-depth research of HDR image quality assessment and convolutional neural network,CNN is used to evaluate HDR image quality in this thesis.The main research contents include:(1)The existing HDR image quality assessment methods based on CNN are improved.The HDR training set is expanded by image transfer and block method,and CNN network is applied to extract the feature vectors from HDR image blocks,which are fused to predict image score using support vector regression.Experimental results show that the improved algorithm can effectively improve the effect of HDR image quality assessment.(2)In order to reduce the demand of the model for the amount of data in HDR image data sets and improve the extraction efficiency of HDR image features by using CNN,an HDR image quality assessment method based on transfer learning and bilateral filtering is proposed in this thesis.This method firstly pre-trains the double-channel CNN network on the SDR training set in order to augment training data.Then the bilateral filtering technology is used to extract the high-brightness salient features of the HDR image during image preprocessing.Finally,combined with the double-channel network,the network is trained using the filtered compressed image and brightness salient feature image to evaluate HDR images’ quality.Experimental results show that this method has a high accuracy rate.