A Local Metric for Defocus Blur Detection Based on CNN Feature Learning
Abstract: Defocus blur detection is an important and challenging task in computer vision and digital imaging fields. Previous work on defocus blur detection has put a lot of effort into designing local sharpness metric maps. This paper presents a simple yet effective method to automatically obtain the local metric map for defocus blur detection, which based on the feature learning of multiple convolutional neural networks (ConvNets). The ConvNets automatically learn the most locally relevant features at the super-pixel level of the image in a supervised manner. By extracting convolution kernels from the trained neural network structures and processing it with principal component analysis, we can automatically obtain the local sharpness metric by reshaping the principal component vector. Meanwhile, an effective iterative updating mechanism is proposed to refine the defocus blur detection result from course to fine by exploiting the intrinsic peculiarity of the hyperbolic tangent function. The experimental results demonstrate that our proposed method consistently performed better than previous state-of-the-art methods. Existing system:
Although the above-mentioned algorithms can be successfully used for image blur region detection, all the existing methods have challenges in precisely differentiating an infocus smooth region and a blurred smooth region. And some proposed blur detection methods still have some problems, e.g., the detection accuracy is not high, the detection time is long, and it is difficult to design blur detector. Although the above-mentioned algorithms can be successfully used for image blur region detection, all the existing methods have challenges in precisely differentiating an infocus smooth region and a blurred smooth region. And some proposed blur detection methods still have some problems, e.g., the detection accuracy is not high, the detection time is long, and it is difficult to design blur detector. Proposed system: However, some proposed blur detection methods still have some problems, e.g., the detection accuracy is not high, the detection time is long, and it is difficult to design blur detector. The more detail information is introduced in section. automatically obtain the local metric map for defocus blur detection, which based on the CNN feature learning. This method no longer requires laborious design of local metrics, and no longer need to know any prior information of the defocus image. We proposed a novel iterative updating mechanism to refine the defocus blur detection result from coarse to fine by exploiting the intrinsic peculiarity of the hyperbolic tangent function. Meanwhile, the experimental results show that. proposed method achieves the state-of-the-art performance and balances the detection accuracy and the detection time powerfully. Advantages: Currently, deep learning has achieved success in different areas of computer vision without relying on prior knowledge, such as Automatic Shadow detection , Saliency detection , Semantic segmentation . Moreover, the local metric is the simple, convenient yet effective method for defocus blur detection. Hence, we will combine the advantages of both ConvNets and local metric to detect defocus blur. In this paper, we proposed a simple yet powerful method.
In this paper, we will combine the advantages of both ConvNets and local metric to detect defocus blur. We have naturally balances the detection accuracy and the detection time powerfully. The architecture of ConvNets is similar to VGG [45]. The best combination of performance parameters . Our method is also efficient. There is a major part in our implementation that uses the most computation time, i.e., the iterative sharpness metric detection by applying the hyperbolic tangent function, but there are also certain speed advantages compared with other methods. A run-time comparison of the different blur detection algorithms is shown in Table . Disadvantages: There have been many attempts in the past two decades to solve the image deblurring problem. Amongst these, it is commonly assumed that the blur kernel is spatially uniform, which allows it to be estimated from global image evidence. Most of those algorithm can efficiently split the blur image into blurred and nonblurred regions by the local metrics of image sharpness. Although the above-mentioned algorithms can be successfully used for image blur region detection, all the existing methods have challenges in precisely differentiating an infocus smooth region and a blurred smooth region. And some proposed blur detection methods still have some problems, e.g., the detection accuracy is not high, the detection time is long, and it is difficult to design blur detector. The local metrics like a filter function to filtering the In this paper, we have proposed a novel yet effective algorithm to address the challenging problem of defocus blur detection from a single image by applying a local sharpness metric, which obtains from the CNN-based of feature learning in the blur and non-blur image regions. Modules: Defocus flour: DEFOCUS blur is an extremely common phenomenon in digital images and is the result of an out-of-focus optical imaging system. Every optical imaging system has a limited depth of field (DOF). The DOF refers to the distance around the image
plane for which the camera is focused. When the camera focuses on the object plane in the image formation process and background is outside that plane or beyond the DOF distance, defocus blur occurs in the resulting image. In digital photography, defocus blur plays an important role in selecting relevant scene information. It can directly attract the attention of the viewer and can emphasize the main subject by making the foreground and background blurry. However, the blurry background restricts the detailed information of the scene, which may suppress computational image understanding and scene interpretation. Hence, blur algorithms are applied to detect the partially blurry image so that postprocessing or restoration algorithms can be applied. Feature learning: we proposed a simple yet powerful method to automatically obtain the local metric map for defocus blur detection, which based on the CNN feature learning. This method no longer requires laborious design of local metrics, and no longer need to know any prior information of the defocus image. The ConvNets architecture used for feature learning consists of alternating multi-convolution and sub-sampling layers, which are similar to VGG architecture . There are numerous convolutional kernels to connect the multi-convolution layer with the sampling layer. The ConvNets architecture extracts the feature information by convolving with each convolutional layer. The output layer is a logistic regression layer that provides a distribution over the classes. In this paper, the output layer includes just two classes, i.e., for the blurry and sharpness feature learning. we have proposed a novel yet effective algorithm to address the challenging problem of defocus blur detection from a single image by applying a local sharpness metric, which obtains from the CNN-based of feature learning in the blur and non-blur image regions. Local sharpness feature: Local sharpness features, e.g. gradient histogram span, kurtosis, for training of a Bayes classifier for blur classification of local image regions . The sharpness is interpreted as the likelihood of being classified as sharp patch. However, the homogeneous regions in the image are weak points in their algorithm. Previous work on defocus blur detection has put a lot of effort into designing local sharpness metric maps or blur detection detector. However, there are a few questions: Whether there are some image features determine the properties of the local area of the image. By extracting convolution kernels from trained neural network structures and processing these convolution kernels with principal
component analysis, we acquire the principal feature, in other words, the local sharpness metric map or blur detection detector. Convnerts: There are numerous convolutional kernels to connect the multi-convolution layer with the sampling layer. The ConvNets architecture extracts the feature information by convolving with each convolutional layer. Zero padding was used whenever size preserving was needed in the learned layers. we have increased the number of convolutional layers in the deep learning network. This multiconvolution layer structure allows ConvNets to learn more hierarchical features. Therefore, ConvNets operate on sets of blurry or sharp patches. We use a batch normalization algorithm to normalize the input data and speed up the training convergence. For the last full connection layer, we chose the CRelu function as the activation function. ConvNets is a supervised learning method. During the training process, we use the backpropagation algorithm to deliver the error of the crossentropy loss function and update the network weight by applying the stochastic gradient descent algorithm. Principal component analysis: Principal component analysis (PCA) is a statistical approach to identify and isolate large amounts of data with multiple variables . In other words, PCA is a featureextraction technique that is capable of distinguishing data values based on their respective variance to the rest of the dataset. We then extract the convolutional kernels from the trained ConvNets architecture and then analyze its principal components. After the convolutional neural network structure is trained, we extract the convolution kernel from it. Each convolution kernel is reshaped into a column, all the convolution kernels are concatenated into a matrix, and then PCA is used to extract the principal components of the matrix. Finally, we reshape the principal component vector that has the maximum explained variance ratio.