Search

Retina vessel segmentation is a vital step in diagnosing ophthalmologic diseases. Traditionally, ophthalmologists segment retina vessels by hand, which is time-consuming and error-prone. Thus, more and more researchers are committed to the research of automatic segmentation algorithms. With the development of convolution neural networks (CNNs), many tasks can be solved by CNNs. In this paper, we propose an encoding-decoding network with a pyramid self-attention module (PSAM) to segment retinal vessels. The network follows a U shape structure, and it comprises stacked feature selection blocks (FSB) and a PSAM. The proposed FSB consists of two convolution blocks with the same weight and a channel-wise attention block. At the head of the network, we apply a PSAM consisting of three parallel self-attention modules to capture long-range dependence of different scales. Due to the power of PSAM and FSB, the performance of the network improves. We have evaluated our model on two public datasets: DRIVE and CHASE_DB1. The results show the performance of our model is better than other methods. The F1, Accuracy, and area under curve (AUC) are 82.21%/80.57%, 95.65%/97.02%, and 98.16%/98.46% on DRIVE and CHASEDB1, respectively.

Micro-expression is a kind of split-second subtle expression which could not be controlled by the autonomic nervous system. Micro-expression indicates that a person is hiding his truly emotion consciously. Because the micro-expression is closely interrelated with lie detection, micro-expression recognition has various potential applications in many domains, such as the public security, the clinical medicine, the investigation and the interrogation. Because recognizing the micro-expression through human observation is very difficult, researchers focus on the automatic micro-expression recognition. This research proposed a novel algorithm for automatic micro-expression recognition which combined a deep multi-task convolutional network for detecting the facial landmarks and a fused deep convolutional network for estimating the optical flow features of the micro-expression. Firstly, this research employed the deep multi-task convolutional network to detect facial landmarks with the manifold related tasks and divided the facial region by utilizing these facial landmarks. Furthermore, a fused convolutional network was applied for extracting the optical flow features from the facial regions which contain the muscle changes when the micro-expression presents. Finally the enhanced optical flow was applied for refining the information of the features and these refined optical flow features were classified by Support Vector Machine classifier for recognizing the micro-expression. The result of experiments on two spontaneous micro-expression database demonstrated that the method proposed in this paper achieved good performance in micro-expression recognition.

Absract ✓Objective: To develop a Convolution Neural Network (CNN) and discuss its performance in the estimation of Glomerular Fifiltration Rate (GFR) for patients with chronic kidney disease (CKD). Methods: A total of 112 patients with chronic kidney disease were enrolled in this study. The GFR is measured by 99mTc-DTPA renal dynamic and used as standard GFR after normalization by body surface area imaging. We established a CNN model and verified the performance of the model by comparing the GFR predicted by the model with the standard GFR. It turned out that the CNN could better evaluate the GFR of patients which is superior to the CG formula, MDRD formula, CKD-EPI formula and GRNN model. Conclusions: The CNN significantly evaluated GFR for patients with CKD, and it showed better performance than traditional methods and GRNN model and closer to the results of 99mTc-DTPA. The experimental results demonstrated that CNN could be used to estimate GFR.

With benefits of low storage costs and high query speeds, binary code representation methods are widely researched for efficiently retrieving large-scale data. In image hashing method, learning hashing function to embed highdimensions feature to Hamming space is a key step for accuracy retrieval. Principal component analysis (PCA) technical is widely used in compact hashing methods, and most these hashing methods adopt PCA projection functions to project the original data into several dimensions of real values, and then each of these projected dimensions is quantized into one bit by thresholding. The variances of different projected dimensions are different, and with real-valued projection produced more quantization error. To avoid the real-valued projection with large quantization error, in this paper we proposed to use Cosine similarity projection for each dimensions, the angle projection can keep the original structure and more compact with the Cosine-valued. We used our method combined the ITQ hashing algorithm, and the extensive experiments on the public CIFAR-10 and Caltech-256 datasets validate the effectiveness of the proposed method.

Objective: Face Recognition is one of the widely application of image processing. Corresponding two-dimensional limitations, such as the pose and illumination changes, to a certain extent restricted its accurate rate and further development. How to overcome the pose and illumination changes and the effects of self-occlusion is the research hotspot and difficulty, also attracting more and more domestic and foreign experts and scholars to study it. 3D face recognition fusing shape and texture descriptors has become a very promising research direction. Method: Our paper presents a 3D point cloud based on mesh local binary pattern grid (Mesh-LBP), then feature extraction for 3D face recognition by fusing shape and texture descriptors. 3D Mesh-LBP not only retains the integrity of the 3D geometry, is also reduces the need for recognition process of normalization steps, because the triangle Mesh-LBP descriptor is calculated on 3D grid. On the other hand, in view of multi-modal consistency in face recognition advantage, construction of LBP can fusing shape and texture information on Triangular Mesh. In this paper, some of the operators used to extract Mesh-LBP, Such as the normal vectors of the triangle each face and vertex, the gaussian curvature, the mean curvature, laplace operator and so on. Conclusion: First, Kinect devices obtain 3D point cloud face, after the pretreatment and normalization, then transform it into triangular grid, grid local binary pattern feature extraction from face key significant parts of face. For each local face, calculate its Mesh-LBP feature with Gaussian curvature, mean curvature laplace operator and so on. Experiments on the our research database, change the method is robust and high recognition accuracy.

In this paper,we address the face sketch recognition problem. Firstly, we utilize the eigenface algorithm to convert a sketch image into a synthesized sketch face image. Subsequently, considering the low-level vision problem in synthesized face sketch image .Super resolution reconstruction algorithm based on CNN(convolutional neural network) is employed to improve the visual effect. To be specific, we uses a lightweight super-resolution structure to learn a residual mapping instead of directly mapping the feature maps from the low-level space to high-level patch representations, which making the networks are easier to optimize and have lower computational complexity. Finally, we adopt LDA(Linear Discriminant Analysis) algorithm to realize face sketch recognition on synthesized face image before super resolution and after respectively. Extensive experiments on the face sketch database(CUFS) from CUHK demonstrate that the recognition rate of SVM(Support Vector Machine) algorithm improves from 65% to 69% and the recognition rate of LDA(Linear Discriminant Analysis) algorithm improves from 69% to 75%.What'more,the synthesized face image after super resolution can not only better describer image details such as hair ,nose and mouth etc, but also improve the recognition accuracy effectively.