Research On The Determination Of Image Distortion Type And No-reference Blur Evaluation

Digital images have different types of distortions during collecting,transmitting,storing and compressing.These types include blur,JPEG compression,Gaussian white noise and so on.Distortion degrades the quality of image.Image quality assessment has become one of the important research directions.A two-phrase assessment system is proposed for image quality assessment: In the first phrase,the specific type of image distortion is determined.In the second phrase,the degree of specific distortion type is assessed.An algorithm based on deep learning is proposed for the first phrase.A deep assessment is made for a specific distortion type of image blur for the second phrase.The main contributions of two phrases are as follows:1.In order to determine type of image distortion,an algorithm based on Gabor wavelet and convolutional neural network(CNN)is proposed.Gabor wavelet is used to extract rough feature of images because of its good characteristic firstly,and then the key feature is extracted from rough feature by improved CNN.The main steps include: 1)Images are preprocessed(Samples is expanded and equalized.And labels are set);2)The preprocessed images are transformed into wavelet domain with 8 directions Gabor wavelet.Eight sub-bands are added to one sample for training;3)A self-designed CNN and Softmax classifier are used to train the final model.The methods of random gradient descent and error back propagation are used to optimize the parameters of convolution kernels during training;4)The final model is used to determine the type of image distortion.The experimental results show that our proposed method can determine various types of image distortion effectively,and has high accuracy and robustness.2.For the specific distortion of image blur,a no-reference assessment algorithm based on discrete cosine transform(DCT)and scale-invariant feature transform(SIFT)is proposed.The algorithm not only makes full use of DCT's frequency decomposition for image blur assessment,but also is closer to human visual system(HVS)by the advantages of SIFT feature points.The algorithm includes the following steps: 1)Grayscale image and gradient map are obtained from the distorted image,then SIFT points are detected in grayscale image;2)The grayscale image and the gradient map are divided into equal-sized blocks respectively;3)The Blocks that contain the feature points in the grayscale image are selected as the interested blocks,the corresponding blocks in the gradient map are transformed into the DCT domain.The sum of squared AC coefficients of DCT(SSAD)is computed;4)In order to eliminate some influence of an image content,the sum of all SSADs is normalized by combination of number of SIFT points,variance and entropy of all interested blocks.The blur score by our algorithm evaluates the degree of image blur accurately.Our blur score is closer to subjective assessment than other methods.