Min Jing

Enhancement of Fiber Orientation Distribution Reconstruction in Diffusion-Weighted Imaging by Single Channel Blind Source Separation

In diffusion-weighted imaging (DWI), reliable fiber tracking results rely on the accurate reconstruction of the fiber orientation distribution function (fODF) in each individual voxel. For high angular resolution diffusion imaging (HARDI), deconvolution-based approaches can reconstruct the complex fODF and have advantages in terms of computational efficiency and no need to estimate the number of distinct fiber populations. However, HARDI-based methods usually require relatively high b-values and a large number of gradient directions to produce good results. Such requirements are not always easy to meet in common clinical studies due to limitations in MRI facilities. Moreover, most of these approaches are sensitive to noise. In this study, we propose a new framework to enhance the performance of the spherical deconvolution (SD) approach in low angular resolution DWI by employing a single channel blind source separation (BSS) technique to decompose the fODF initially estimated by SD such that the desired fODF can be extracted from the noisy background. The results based on numerical simulations and two phantom datasets demonstrate that the proposed method achieves better performance than SD in terms of robustness to noise and variation in b-values. In addition, the results show that the proposed method has the potential to be applied to low angular resolution DWI which is commonly used in clinical studies.

Uncertainty Reasoning Based Formal Framework for Big Video Data Understanding

It is worthwhile to incorporate human knowledge with conventional machine learning approaches for big data analytics. Focusing on big video data understanding, this paper presents a formal scenario recognition framework where knowledge-based logic representation and reasoning is combined with data-based learning approach to enhance scenario recognition capabilities. This is achieved via multi-layered (hierarchical) processing. This approach constructs the hierarchical representation structure based on the semantic understanding of considered scenario, and transforms the structure into logic formulas. After applying conventional computer vision methods for low-level events classification, we apply logic based uncertainty reasoning to determine scene content. Experimental results on a benchmark dataset are provided to show the rationality of the proposed approach.

Integration of text and image analysis for flood event image recognition

Flood event monitoring plays an important role for emergency management. With the fast growth of social media, a large number of images and videos are uploaded and searched on the internet during disasters, which can be used as “sensors” for improving efficiency of emergency management. This work proposes a novel framework in which the rich information available from social media is incorporated with image analysis to enhance image retrieval for disaster management. The text associated with images of flooding events was used to extract prominent words associated with flooding. The image features are represented by a histogram of visual words obtained using the Bag-of-Words (BoW) model. The text and image analysis are integrated at the feature level, in which the text features are conjoined directly with image features. The proposed approach was evaluated based on two flood event corpuses obtained from the US Federal Emergency Management Agency media library and public Facebook pages and groups related to flood and flood aid (in German). The experimental results demonstrate the improved performance of image recognition after incorporating the text features, which suggests the potential to enhance the efficiency of emergency management.

A novel spiral addressing scheme for rectangular images

Spiral architectures have been employed as an efficient addressing scheme in hexagonal image processing (HIP), whereby the image pixel indices can be stored in a one-dimensional vector that enables fast image processing. However, this computational advance of HIP is hindered by the additional time and effort required for conversion of image data to a HIP environment, as existing hardware for image capture and display are based predominantly on traditional rectangular pixels. In this paper, we present a novel spiral image processing framework that develops an efficient spiral addressing scheme for standard square images. We refer to this new framework as “squiral” (square spiral) image processing (SIP). Unlike HIP, conversion to the SIP addressing scheme can be achieved easily using an existing lattice with a Cartesian coordinate system; there is also no need to design special hexagonal image processing operators. Furthermore, we have developed a SIP-based non-overlapping convolution technique by simulating the “eye tremor” phenomenon of the human visual system, which facilitates fast computation. For illustration we have implemented this technique for the purpose of edge detection. The preliminary results demonstrate the efficiency of the SIP framework by comparison with standard 2D convolution and separable 2D convolution.

Temporal Changes of Diffusion Patterns in Mild Traumatic Brain Injury via Group-Based Semi-blind Source Separation

Despite the emerging applications of diffusion tensor imaging (DTI) to mild traumatic brain injury (mTBI), very few investigations have been reported related to temporal changes in quantitative diffusion patterns, which may help to assess recovery from head injury and the long term impact associated with cognitive and behavioral impairments caused by mTBI. Most existing methods are focused on detection of mTBI affected regions rather than quantification of temporal changes following head injury. Furthermore, most methods rely on large data samples as required for statistical analysis and, thus, are less suitable for individual case studies. In this paper, we introduce an approach based on spatial group independent component analysis (GICA), in which the diffusion scalar maps from an individual mTBI subject and the average of a group of controls are arranged according to their data collection time points. In addition, we propose a constrained GICA (CGICA) model by introducing the prior information into the GICA decomposition process, thus taking available knowledge of mTBI into account. The proposed method is evaluated based on DTI data collected from American football players including eight controls and three mTBI subjects (at three time points post injury). The results show that common spatial patterns within the diffusion maps were extracted as spatially independent components (ICs) by GICA. The temporal change of diffusion patterns during recovery is revealed by the time course of the selected IC. The results also demonstrate that the temporal change can be further influenced by incorporating the prior knowledge of mTBI (if available) based on the proposed CGICA model. Although a small sample of mTBI subjects is studied, as a proof of concept, the preliminary results provide promising insight for applications of DTI to study recovery from mTBI and may have potential for individual case studies in practice.

Human Action Recognition in Video via Fused Optical Flow and Moment Features Towards a Hierarchical Approach to Complex Scenario Recognition

This paper explores using motion features for human action recognition in video, as the first step towards hierarchical complex event detection for surveillance and security. We compensate for the low resolution and noise, characteristic of many CCTV modalities, by generating optical flow feature descriptors which view motion vectors as a global representation of the scene as opposed to a set of pixel-wise measurements. Specifically, we combine existing optical flow features with a set of moment-based features which not only capture the orientation of motion within each video scene, but incorporate spatial information regarding the relative locations of directed optical flow magnitudes. Our evaluation, using a benchmark dataset, considers their diagnostic capability when recognizing human actions under varying feature set parameterizations and signal-to-noise ratios. The results show that human actions can be recognized with mean accuracy across all actions of 93.3%. Furthermore, we illustrate that precision degrades less in low signal-to -noise images when our moments-based features are utilized.

A novel criterion for characterizing diffusion anisotropy in HARDI data based on the MDL technique

Based on the spherical harmonic decomposition of HARDI data, we propose a new criterion for characterizing the diffusion anisotropy in a voxel directly from the SH coefficients. Essentially, by considering the Rician noise in diffusion data, we modify the Rissanens criterion for fitting the diffusion situation in a voxel. In addition, the minimum description length (MDL) criterion has been employed for interpreting information from both the SH coefficients and the data. The criterion obtained can make use of the diffusion information so as to efficiently separate the different diffusion distributions. Various synthetic datasets have been used for verifying our method. The experimental results show the performance of the proposed criterion is accurate.

Novel “Squiral” (square spiral) architecture for fast image processing

Abstract Fast image processing is a key element in achieving real-time image and video analysis. We propose a novel framework based on a square spiral (denoted as “squiral”) architecture to facilitate fast image processing. Unlike conventional image pixel addressing schemes, where the pixel indices are based on two-dimensional Cartesian coordinates, the spiral addressing scheme enables the image pixel indices to be stored in a one dimensional vector, thereby accelerating the subsequent processing. We refer to the new framework as “Squiral” Image Processing (SIP). Firstly we introduce the approach for SIP conversion that transforms a standard 2D image to a 1D vector according to the proposed “squiral” architecture. Secondly we propose a non-overlapping convolution technique for SIP-based convolution, in which the SIP addressing scheme is incorporated by simulating the phenomenon of eye tremor in the human visual system. Furthermore, we develop a strategy to extend the SIP framework to be multiscale. The performance of the proposed framework is evaluated by the application of SIP-based approaches to edge and corner detection. The results demonstrate the efficiency of the proposed SIP framework compared with standard 2D convolution.

BIOLOGICALLY MOTIVATED SPIRAL ARCHITECTURE FOR FAST VIDEO PROCESSING

Fast image processing is a key element in achieving real-time image and video analysis. The spiral addressing scheme [10] has been an efficient tool for hexagonal image processing (HIP), whereby the image pixel indices are stored in a one-dimensional vector that enables fast processing. Unlike HIP, which requires a complex resampling scheme, we present a novel “squiral” (square spiral) image processing (SIP) framework that provides a spiral addressing scheme for direct application to standard square pixel-based images. A SIP-based non-overlapping convolution technique is developed by simulating the eye tremor phenomenon of the human visual system to accelerate computation in feature extraction. Furthermore, we deploy the proposed simulated eye tremor technique on a sequence of video frames. The preliminary results based on two action video clips demonstrate the potential of the SIP-based eye tremor model to facilitate fast video processing.

Recognition by Enhanced Bag of Words Model via Topographic ICA

The Bag-of-Words (BoW) model has been increasingly applied in the field of computer vision, in which the local features are first mapped to a codebook produced by clustering method and then represented by histogram of the words. One of drawbacks in BoW model is that the orderless histogram ignores the valuable spatial relationships among the features. In this study, we propose a novel framework based on a topographic independent component analysis (TICA), which enables the geometrically nearby feature components to be grouped together thereby bridge the semantic gap in BoW model. In addition, the compact feature obtained from TICA helps to build an efficient codebook. Furthermore, we introduce a new closeness measurement based on Neighbourhood Counting Measure (NCM) to improve the k Nearest Neighbour classification. The preliminary results based on KTH and Trecvid data demonstrate the proposed TICA/NCM approach increases the recognition accuracy and improve the efficiency of BoW model.