Blur Image Quality Assessment And Detection Based On Feature Learning

Digital image has become an indispensable and key information carrier in the era of artificial intelligence and big data.The Full life cycle of digital image,such as acquisition,transmission,processing,storage,etc.,will be inevitably accompanied by the decline of image visual quality and the distortion of intrinsic information.Blur is one of the most common forms of digital image degradation and distortion.Blur not only affects the visual perception of the image intuitively,but also leads to a loss of important knowledge contained in an image.Therefore,it is necessary to evaluate the image quality precisely before the image reaches the terminal.Accurate detection and segmentation of the local blur region is the primary prerequisite for targeted image restoration,and is particularly important for mining and utilizing the deep information contained in the scene.Therefore,it is an important and challenging topic to evaluate the overall quality effectively and detect the local blur of the blur image accurately in the field of computer vision.This paper does research on the blur image and probes into the mining,extracting and learning of blur features in frequency domain,sequency domain,deep level,multi-scale and complex scenes.Results and innovations of this paper mainly include four aspects as following:1.Aiming at global blur,an objective quality assessment algorithm for no-reference no-training frequency-domain-feature based image blurriness estimation(NRNT-FIBE)is proposed.By combining the re-blur strategy and discrete cosine transform,the NRNT-FIBE algorithm can effectively quantify the quality of global blur images in defocus,motion,shake and other blur forms.At the same time,the frequency domain features constructed based on discrete cosine transform are concise.Compared with other no-referenced blur image quality assessment algorithms based on handcrafted features,the no-trained NRNT-FIBE algorithm can quickly characterize and calculate the blurriness of the image,which has advantages in computing efficiency.2.Aiming at global blur,an image quality classification algorithm based on depth dense features(DN-BIQC)is proposed.It can accomplish the fast binary classification for Blur/Sharp of digital images,and can also classify the images into five categories for Excellent,Good,Fair,Poor and Bad according to their blurriness,so as to accurately filtrate and separate images with different blurriness.The algorithm applies the densely connected convolutional neural network to extract deep dense features,which helps to overcome the vanishing gradient problem caused by the increasing depth of the convolutional neural network,and thus improves the feature mining capability and feature utilizing efficiency of the network.Experiment results show that the DN-BIQC algorithm has high classification accuracy and strong robustness against noise.3.Aiming at local blur,a fast algorithm for no-reference no-training sequency-domain-feature based blur detection and segmentation(NRNT-SBDS)is proposed.Combined with the re-blur strategy and the Walsh-Hadamard transform,a feature that can effectively measure the blurriness of local blur image by pixels is explored in the sequency domain to achieve the fast detection of local blur image.Finally,by combining K-means clustering and morphological operation,the local blur region of the image can be segmented accurately.4.Aiming at local blur,an algorithm for complex blur detection based on multi-scale fusion features and pyramid M-shaped ensemble model(PM-Net)is proposed.A new multi-input multi-loss encoder-decoder network(M-net)is developed to study the blur features of different scales from coarse-to-fine.Then,a pyramid ensemble model(PM-Net)composed of multi-scale M-nets and a unified fusion layer is proposed to solve the scale ambiguity problem.PM-Net can achieve the accurate detection of anomalous regions(such as homogeneous but clear regions and pseudo-clear backgrounds),multi-form blur scenes and other complex blur regions.Meanwhile,it also can overcome the problem of scale ambiguity problem.Noted that PM-Net(millisecond level)is hundreds of times faster than other state-of-the-art methods(second level),and has strong noise robustness and good model generalization ability.