Image Quality Assessment Based On Deep Learning

The rapid development of Internet leads to a dramatic increase in the number of images.Image has played an important role in our life.However images can't avoid becoming poor in terms of quality when we get or transfer them.The quality of images has an important influence on getting information and experience when we look through the images.Image quality assessment(IQA)algorithms aim at designing an efficient model to predict the image quality automatically and precisely,so that we can get good images from quantity of images and optimize the image processing algorithms and systems,and finally we can improve the experience when we look through images.So it is meaningful to develop IQA algorithms.We studied IQA in terms of fidelity and usability of image.Our works are as follow:Firstly,in the aspect of fidelity,we proposed a novel full-reference(FR)IQA framework based on the similarity of deep features,codenamed deep similarity(Deep Sim).The existing IQA algorithms are based on handcraft feature or shallow neural network.Using handcraft feature needs professional knowledge and can hardly get good results.Shallow neural network is limited in image representation.Due to the success of the convolutional neural networks(CNNs),we proposed a novel full-reference(FR)IQA framework based on the similarity of deep features.We extract features from reference images and test images respectively and calculate the local similarity of every layer of feature maps which are regarded as a form of image quality.Finally we synthesize local quality to get the overall image quality.The precision of Deep Sim is as high as 90%.Compared with the exist methods,the overall performance of Deep Sim on four datasets has an improvement of above 1%.We also investigate the effect of the depth and layer type on the IQA performance and analyze how the performance of Deep Sim varies with the pooling strategies.In the aspect of usability,we investigated the effect of image quality on the prediction for the intestinal hemorrhage detection.Inspired by the conclusion of the Deep Sim,we augmented the dataset by changing the quality of images through rotation,blurring,luminance change and Poission noise and trained four classic CNNs with the augmented dataset.Then we compared the results that we get with augmentated dataset and original dataset.We predicted different type and class distorted images with our model to investigate the effect of different distortion on prediction.We dramatically improved the precision from 45.67% to 98.47% and avoided overfitting and we found that luminance has the greatest impact on the result.All in all,our work has some theoretical innovation and we made a breakthrough in the result.And our work is beneficial to understand the image quality and generalize the related technology.