ABSTRACT. Decomposing discrete signals such as images into components is vital in many applications, and this paper propose a framework to produce filtering banks to accomplish this task. The framework is an equation set which is ill-posed, and thus have many solutions. Each solution can form a filtering bank consisting of two decomposition filters, and two reconstruction filters. Especially, many existing discrete wavelet filtering banks are special cases of the framework, and thus the framework actually unified different wavelet filtering banks. Moreover, additional constraints can impose on the framework to make it well-posed, meaning that decomposition and reconstruction (D&R) can consider the practical requirements, not like existing discrete wavelet filtering banks whose coefficients are fixed. All the filtering banks produced by the framework can behave excellently, have many decomposition effect and precise reconstruction accuracy, and this has been theoretically proved and been confirmed by a large number experimental results.

ABSTRACT. The segmentation of images with intensity inhomogeneity is always a chal-lenging problem. Due to the presence of inhomogeneous intensities, traditional intensity-based segmentation algorithms may lead to serious misclassification. In this paper, unlike some of the existing algorithms that use local region descriptors to reduce or correct the intensity inhomogeneity, we present a framework which can make use of it to help to accomplish segmentation tasks. Our main argumen-t is to quantify in- tensity inhomogeneity and convert it to be a useful feature to improve the segmentation results. We formulate a level set segmentation frame-work based on the quantification of intensity inhomogeneity. Two new quantities called the region intensity inhomogeneity index and factor are firstly introduced to describe the degree of intensity inhomogeneity of images. Then the quantified intensity inhomogeneity is incorporated into a variational level set framework. The total energy functional of the proposed framework consists of a region intensity inhomogeneity term, a local region fitting energy, and an edge based energy. The first term is defined by exploiting the region intensity inhomogeneity factor of the images. Mean and variance of the image intensity are utilized in the local region fitting energy to describe the texture information in the images. The last term of the energy functional, which is also derived from the region intensity inhomogene-ity factor, incorporates the significant edge information. By integrating these three terms, the proposed model can handle images with extremely inhomogeneous in-tensities, especially two kinds of images: one with homogeneous intensities in the objects and inhomogeneous intensities in the backgrounds, another one with homo-geneous intensities in backgrounds and inhomogeneous intensities in the objects. In this way, the intensity inhomogeneity is converted into usable information by introducing the quantification of it into the energy functional. As far as we know, it is the first attempt to quantify the intensity inhomogeneity of images and convert it into useful information to aid the segmentation. Experimental results show that the proposed model is robust to initializations. Results on our collected 40 images with the two types of inhomogeneities mentioned above and ECSSDs 1000 natural images indicate that the proposed algorithm can produce satisfactory segmentation-s and outperform in comparison to five state-of-art methods. This is a joint work with Xu Li and Hairong Liu

ABSTRACT. Magnetic particle imaging (MPI) is a tracer-based imaging modality developed to detect the concentration of superparamagnetic iron oxide nanoparticles. It is highly sensitive to the nanoparticle's nonlinear response to the applied magnetic field. Highly dynamic applied magnetic fields allow a rapid data acquisition in 3D. Model-based reconstruction techniques are still not able to reach the quality of data-based approaches in which the linear system function is determined by a time-consuming calibration process. Possible reasons include the behavior of nanoparticles in the rapidly changing magnetic fields. The simple model of paramagnetism described by the Langevin function (equilibrium model) is not sufficient to describe emerging relaxation-like effects. More sophisticated models are required to gain reconstructions with comparable quality. In this talk we discuss the model-based reconstruction problem in magnetic particle imaging and how magnetization dynamics influence the system behavior and thus the concentration reconstruction. We further analyze the inverse problem obtained from the equilibrium model.

ABSTRACT. Terahertz radiation has great application potentials for security inspection due to its unique properties of good penetrating ability and harmlessness for human body. Terahertz imaging is a promising technique for the remote detection of hidden objects concealed on human body. This presentation will provide an overview of main features and recent advances of active and passive terahertz imaging techniques for personnel screening.

ABSTRACT. Arterial luminal stenosis in cerebral artery can result in insufficient blood perfusion in the brain. However, the severity of stenosis shown by medical images does not necessarily reflect the underlying hemodynamic alteration caused by the vessel narrowing. The fractional pressure ratio (FPR) is therefore introduced to quantitatively assess the hemodynamics significance of severe intracranial stenosis. A computational fluid dynamics (CFD) based method using high performance computing is proposed to non-invasively compute the FPRCFD and compared against FPR measured by an invasive technique. The proposed CFD method can potentially be useful in assessing the functional alteration of cerebral stenosis.

ABSTRACT. Neurons have been considered the most important functional unit of the brain and have attracted the most attention in neurobiology. Neural cells occupy 70% - 80% of the whole brain volume, however, our current knowledge of neural circuit can only partially explain the neurological and psychiatric conditions of the brain. Thus, it is also important to consider the influence of the Interstitial system (ISS) around neural cells, which is consisted by extracellular space（ECS）and the interstitial fluid (ISF) and plays crucial roles in substance transport and signal transmission among neurons. Brain ECS is the major compartment of the brain microenvironment that occupies 15% to 20% of the total brain volume. It provides the immediate accommodation space for neural cells and is a complex and dynamic space connecting the vascular system and neural networks. This paper includes the anatomy of the brain ECS under physiological condition, the in vivo measurement techniques of brain ECS, and the divisions of brain ECS. Moreover, the implications of ECS knowledge of basic neuroscience and clinical applications are introduced, including the brain tissue engineering and the local drug delivery via ECS.

ABSTRACT. In recent years, the UAV technology has developed rapidly and played an important role in many fields, especially in intelligence, reconnaissance and monitoring. Object detection can provide accurate target location and target category for reconnaissance missions, providing detailed command information for commanders. However, the current object detection algorithm based on deep convolution network does not well on detection for small objects and so cannot be well applied to small objects in the reconnaissance image of UAV. In this paper, an object detection algorithm for UAV reconnaissance image based on deep convolution network is proposed. The image is adaptively divided according to the UAV flight parameters and the payload parameters before sent into the network. Through this way, small objects can be located and classified in a high accuracy of location and classification. This method can detect objects with small size, multiple quantities and multiple categories on UAV.

ABSTRACT. The detection of metal edges is a common problem in manufacturing of metal products including strips and tapes, as well as in alignment of metal sheets for assembly of products. The detection may be as simple as detecting the position of the edge – either qualitatively, that is, the very existence of a metal edge or may be highly quantitative, whereby the exact position of the edge is sought for the purpose of alignment or assembly. Metal strips may be very thin or relatively thick, may be ferromagnetic or nonmagnetic and in some applications, such as during production, may be at high temperatures. The present work proposes a self-compensating, highly sensitive resonant sensor suitable for the industrial environment. The method relies on two split coils one for each edge, detecting the position of one edge of the strip. To compensate for a variety of effects such as jitter, drift and frequency locking, the coils are switched at a fixed rate so that only one circuit resonates at any given time and the difference in resonant frequency of the two resonators is used to locate the edges and to generate correction signals for the purpose of positioning the metal strip. The results presented here are based on a prototype system.

ABSTRACT. Weak magnetic stress internal detection technology of the pipeline can effectively detect the stress concentration regions where the pipeline has not formed macroscopic defects, and advance the pipeline safety alarm time, which makes up for the weaknesses of the traditional magnetic flux leakage internal detection technology. In this paper, the key technologies of the weak magnetic stress internal detection are stated. The detection mechanism, influencing factors and engineering tests of weak magnetic stress detection technology are deeply analyzed. Furthermore, the future development direction of the weak magnetic stress internal detection technology is proposed.

ABSTRACT. Compared with conventional computed tomography (CT), dual-spectral comput-ed tomography (DSCT) has advantages for material differentiation, artifact re-duction and so on. In the DSCT, two sets of projections from two different spec-tra are essential. However, available commercial DSCT equipments have consid-erable hardware effort, limiting their applications. Here we present a dynamic fil-ter design that allows acquiring low/high-spectral projections with a full single CT scan. We also propose a corresponding iterative reconstruction method. The vir-tual monochromatic image is linearly composed of two basis material images re-constructed by the proposed method from the acquired DSCT data set. The re-sults of the simulated dental CT experiments suggest the effectiveness of the pro-posed method in eliminating metal artifacts in monochromatic images.

ABSTRACT. This paper proposes an online geometric calibration method on cone beam CT(CBCT), based on Epipolar geometry consistency. Our method is a new online approach with high calibration accuracy, comparable to that of offline one. Four parameters: the detector skew, rotation axis, mid-plane and the source to detector distance are served as modeling the geometry of CBCT in this paper. An objective function is deduced by exploiting the Epipolar geometry consistency among the projective views and the optimal parameters are obtained by taking advantage of SIMPSA algorithm to minimize the objective function. In the simulations, different noise levels are added respectively and the results show that our method is insensitive to noise. In the real experiment, the bamboo stick and Chinese parasol tree branch imaging manifests that the accuracy of our method is comparable to that of the offline calibration ones.

ABSTRACT. This paper focuses on moving target detection in single channel Circular-SAR(CSAR). The logarithm background subtraction algorithm described in this paper utilizes the overlapped subaperture logarithm image sequence to detect moving targets. It first model the background image with the input image se-quence. Then using the input images subtract background image to cancel the clut-ter. Finally the moving target can be detected in subtracted image sequence. How-ever the detection performance depends on the azimuth aperture width(otherwise number of input images in one set of image sequence). Thus the detection perfor-mance is analyzed with two measurements: signal-to-clutter noise ratio(SCNR) improvement and clutter cancellation ability. Based on the analysis, the proper az-imuth aperture width for achieving best detection performance could be obtained. The algorithm is validated with GOTCHA-GMTI dataset.

ABSTRACT. In this work, we investigate a novel regularization term on dual-energy CT (DECT) limited-angle reconstruction problem. To better suppress the artifacts and noises caused by limited-angle scan and dual-energy decomposition, as well as preserve the edges of scanned object, we present a regularization term called WDL0+TV. We solve the WDL0+TV based DECT reconstruction model with iteration procedure and sequence the regularization term in iterative steps. Then, a two-step solver is proposed to numerically solve this problem. Experimental results indicate that the reconstructed images can be visually comparable to the images from full angle.

ABSTRACT. Fuzzy clustering analysis algorithm has good ability to solve fuzzy problems. Rough clustering analysis has good ability to solve prior knowledge uncertain problems. But in the real world, there are many problems that not only are fuzzy but also are rough and uncertainty, the paper combines the idea of these two algorithms. In order to improve correction of clustering result, it imports attributes reduction algorithm to get importance of each attribute, and dynamically changes attribute weight by its importance. The new algorithm firstly computes fuzzy membership degree of every object, and then estimates the object belong to lower approximation or upper approximation of one cluster. In the analysis process, the paper provides a new way to get the cluster centers, combing fuzzy and rough theory. From experiments for four UCI datasets, it proves the new algorithm is better effective.

ABSTRACT. The projection data can be modeled as a piecewise smooth function. It’s well known that L1 regularization tries to recover piecewise constant functions, while L2 aims to recover smooth functions. This motivates us to propose the Lp regularization with 1 < p < 2 for low-dose projection data denoising. The resulting Lp regularized model is then linearized such that the fast split-Bregman algorithm can be employed. Experiments on simulated projection data show that Lp regularization with 1 < p < 2 could achieve better reconstruction compared to L1 regularization.

ABSTRACT. As a kind of emerging weapon, UA V (Unmanned Aerial Vehicle) has become increasingly prominent in modern warfare and received unprecedented attention from the world ’s military powers. The major function of UAV in warfare is reconnaissance, providing intelligence support for systematic warfare with the equipped different kinds of payloads. Videos and images are the most intuitive reconnaissance results and the basic carrier of the reconnaissance information processing system. Th is paper analyzes the application of image processing technology in UAV intelligence information processing through requirements analysis of reconnaissance missions-oriented and teasing the reconnaissance mission flow of UAV. We finally complete the design and implementation of UAV reconnaissance information processing system.

ABSTRACT. Synthetic aperture imaging is widely applied to removing occlusion. Unfortunately, due to the influence of occlusion, the quality of reconstruction result is undesirable. Therefore, how to distinguish foreground and background and screening effective rays become the key to improve the quality of reconstructed image. In this paper, we propose a method to remove interference from occlusion before reconstructing objects. By focusing on the foreground occlusion, occlusion plane will be sharp and other planes will be blurred. Taking advantage of this strong contrast, it is easy to segment occlusion and background when we apply image matting. The quality of reconstruction image will be improved when synthetic aperture imaging without occlusion. We demonstrate the superiority of our method by presenting experimental results as well as comparing our method with tradition approach.

ABSTRACT. Lung nodule detection plays an important role in the early diagnosis of lung cancer. Traditional lung nodule detection methods are complex and time-consuming. In order to improve lung nodule detection efficiency and help physicians to diagnose early and accurately, we proposes a novel computer-aided detection method for lung nodule detection in CT images using multi-level feature fusion convolutional neural network, which motivated by the work of Faster R-CNN. Firstly, the convolutional neural network extracts features from training data automatically and merges multi-level features to obtain fused feature maps with rich and discriminative information, then fully convolutional neural network is used to generate candidate nodules on the fused feature maps, finally nodule classification and regression is performed on candidates. On 888 scans of the publicly available LIDC-IDRI dataset, our method reaches high detection sensitivities of 85.6% and 90.3% at 1 and 4 false positives per scan, respectively. The proposed method improves the detection rate of lung nodules, and physicians could diagnose lung lesions effectively under the help of this method, it is important to the diagnosis of lung cancer.

ABSTRACT. In the process of behavior recognition of multi-players for soccer game video, various features of athletes need to be extracted. In this paper, color moments extracted by using color classification learning set are regarded as color feature. Contour features of athletes are extracted by utilizing players silhouettes block extraction and normalization. Hough transform is used to extract the features of coordinates of pitch line, which can be used for camera calibration, rebuilding the stadium and calculating the coordinate of players in the real scene. The trajectories of players and ball are predicted by using Kalman filter, while trajectories characteristics of player and ball are extracted by using the trajectory growth method. Experimental results show that the accuracy of behavior recognition can be greatly improved when these features extracted are used to recognize athlete behavior.

ABSTRACT. Multi-scale decomposition (MSD) strategy is wildly applied to extract characteristics at different scales in image fusion field, and then the different image components will stay in different layers. To avoid the negatively impact that using the same fusion rule on different scales, a Gaussian filter MSD-based of fusion method with total variation (TV) and phase congruency (PC) is proposed in this paper. This framework consists of three parts: First, decompose the source images into base layer and detail layer through Gaussian filter; Then, TV and PC are set as the representations of base and detail layers, respectively; Finally, combine the fusing two layers to reconstruct the fused image. Some fusion methods and several quantitative analyses are displayed, and the experimental results show that the proposed method has a superior performance compared with some other popular fusion methods, both in human vision and objective evaluations.

ABSTRACT. The point cloud data collected by the 3D scanner can accurately represent the surface information of the object, and Computed Tomography (CT) technology reconstructs internal density information through projection of the object. In order to observe internal and external information altogether, this paper proposes a new method for the registration of point cloud data and 3D CT image. Firstly, the coarse registration is achieved with point cloud registration which introduces the probability threshold to improve the accuracy, then the fine registration is obtained by voxel registration through the surface density and gradient information of CT image. The results of experiment show that the proposed method is superior in robustness, accuracy and efficiency.

ABSTRACT. Hyperspectral images (HSIs) contain abundant spectral and spatial in- formation which is very useful for various applications. Since the data structure is highly complex, the key problem is how to extract useful information we want. Traditional feature extraction methods go the way, however, these methods un- dergo severe limitations. Most of them are designed mathematically not physically and ignore the fact that the changes of physical imaging conditions will make a significant influence in the spectra obtained. In this paper, we try to analysis the information contained in HSIs from the perspective of hyperspectral imaging prin- ciple, and propose a novel method for extracting reflectance and surface normals from HSIs.