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

Nowadays,with the popularity of products such as smart phones,images have become the main form of transmitting of information.It has made the rapid growth of mobile data.It may cause a certain degree of distortion in the process of image transmission,so it needs a good way to effectively evaluate the image quality.In the actual life,it is difficult to obtain the original reference image of the distorted image.Therefore,the research content of this paper is image quality evaluation without references based on the deep learning method.Deep learning has developed rapidly in the latest years,especially in the fields of image processing,speech recognition.It has been well applied and has achieved good results.Deep learning can provide sufficient training for the data and can learn features well.The trained model can reflect the potential information of the data.So this paper applies this to image quality assessment.The content of the paper is mainly described as follows:1.We evaluate the mixed distorted image quality without reference based on deep belief networks.This method first normalizes the local brightness coefficient of the image in the mixed distortion image database,and then cuts into 20×20 non-overlapped images.In the image block,the small block matrix is reconverted into a 1×400 matrix,and one row of the matrix represents one small block.Afterwards,according to the database DMOS value range,it is divided into twenty categories,and each small block is classified according to its DMOS value,a small block label is made,and a training set and a test set are completed.Then input it into the DBN model to train and learn,and get the classification of the image quality class.Finally,the quality of the hybrid distorted image is obtained by the designed quality pooling method,and the evaluation index for the entire LIVE mixed distortion database is obtained.2.A method that can predict stereoscopic image quality quickly based on Siamese network was constructed.With this method,the normalized stereoscopic image pair is used as the input of the Siamese network.Using the similarity of the left and right views and the Siamese network principle,the weights in the left-view learning process are used as the weights in the right-view learning process.Reduce the amount of calculation and calculation time.The Siamese network has two branches and they share weights with each other.And then it uses the contrastive loss function to enhance the ability of network to identify data sets.Finally,a regression model can be used to obtain the predicted quality scores of the stereoscopic images.3.According to image fusion retaining the information of original images and reducing calculation,an algorithm of stereoscopic image quality assessment based on image fusion and CNN is proposed.The left view and the right view are fused by the PCA fusion method.Then the localized average brightness is subtracted from the fused image and the contrast is normalized.Then 40×40 non-overlapping small blocks are selected as input for each image.The convolutional neural network trained an evaluation model that can reflect the mapping relationship between the stereoscopic image quality features and the subjective score differences.The scores of the stereoscopic image patches were obtained.The average value of each stereoscopic image was the predicted quality score of the entire distorted stereoscopic image.The final experiment shows that this method has obtained better evaluation results.4.Based on a binocular vision,a deep convolutional neural network is designed and the method of multi-channel inputs is used to evaluate the stereoscopic image quality.The inputs of the two channels are left and right views.The three-channel input uses the left and right views and the disparity map as input to the convolutional neural network.The disparity map contains depth information of stereoscopic images.Then convolutional layer and pooling layer is performed for each channel.The last pooling uses the joint of maximum pooling and median pooling.Finally,the eigenvectors output from the three channels are linearly spliced and they are input into the fully connected layer together.At last,we can obtain the regression model between image quality characteristics and scores.Finally,experiments were performed on the LIVE 3D image database.The results reflect that the method is quite good.