Search

Here, we overview an optical coherence tomography image-guided needle puncture based on recent results, which can help operators more accurately indicate the depth position of the puncture needle into the body, reduce subjective judgment errors, and assist the physician in operating a smoother procedure, and increase the success rate of the operation. Artificial intelligence aided image interpretation can further provide an automatic diagnosis in the future.

A tethered viewpoint in an operator interface visualizes remote situations according to a robot’s heading direction. The tethered view, adhering to the line-of-sight requirement, results lower visual momentum when visualizing dense environments, such as during indoor teleoperation tasks. This problem occurs because the tethered view only partially visualizes spatial information, unlike the bird’s-eye view. Operators are thus inhibited to build their spatial mental model because the transitions happen as the robot moves. This paper presents an approach to improve the visual momentum of the tethered view by complementing the omitted spatial information. The approach augments simplified spatial cues on the excluded areas of a tethered view. The cues are purportedly used to illuminate the basic spatial structure of the surroundings and can thus help operators in constructing their spatial mental models. Evaluation of the presented approach was conducted using a simulated telerobot environment. The results indicated that the augmented view possessed a higher visual momentum, which exhibited by the lowered workload and enhanced levels of both spatial perception and situational awareness.

The presence of 3-D maps of remote environments offers realistic visualization of both front and surrounding views. The surrounding viewallows operators to observe different side views by controlling their viewpoint. Viewpoint control can be achieved by using the orientation of the operators’ input device, which is mostly akin to gamepad controller. This kind of method assumes that binding the gamepad orientation to the 3-D map could benefit the operators during tasks; therefore, assessing operators’ performance is required to validate this assumption. This paper assessed the effect of using gamepad orientation for controlling 3-D map viewpoint during remote navigation. The evaluation was conducted by comparing a tethered viewpoint against a controllable viewpoint within a simulated telerobot situation. The evaluation resulted in a contradictive state, where participants’ positive opinions and preferences were against their actual workloads ( ${d}$ = 0.299, 95% CI [−0.200, 0.792]). The contradiction indicated the presence of naive realism on using gamepad orientation for rapidly controlling viewpoint of 3-D map views. Meanwhile, shifts in participants’ navigation behavior were also observed.

A point of interest (POI) is a specific location that people may find useful or interesting, such as restaurants, stores, attractions, and hotels. With the recent proliferation of location-based social networks (LBSN), numerous users gather to interact and share information on various POIs. POI recommendations have become a crucial issue because it not only helps users to learn about new places but also gives LBSN providers chances to post POI advertisements. As we utilize a heterogeneous information network to represent an LBSN in this work, POI recommendations are remodeled as a link prediction problem, which is significant in the field of social network analysis. Moreover, we propose to utilize the meta-path-based approach to extract implicit but potentially useful relationships between a user and a POI. Then, the extracted topological features are used to construct a prediction model with appropriate data classification techniques. In our experiments, the Yelp dataset is utilized as our testbed for performance evaluation purposes. The results show that our prediction model is of good prediction quality in practical applications.

This paper presents a novel video stitching system for stitching multiple images with 4K and displaying them simultaneously. The key insight of the proposed system can be divided into two stages First, a image of super resolution, such as 16K/32K, and it can be divided evenly into 2×2 or 3×3…etc. Due to the parallelized GPU, we could decode these image simultaneously based on the DXT and JPEG compression. However, the next key point is how to synchronize the split image of vide for display. After image decoding, the synchronized queue buffer are implemented to store the decompressed 4K images for stitching synchronized video frames. The experimental results demonstrate the stitched seamless videos with four/eight times the 4K resolution can be displayed on the physical electronic display wall in real time, and the frame rate of the resultant videos achieves 30–60 frames per second (FPS). Consequently, the proposed novel video stitching system provide satisfied performance of video watch for the future generation of 8K/16K video processing systems.

Identifying a specific user is an old but challenging problem, and its applications are ubiquitous in our daily lives. Conventional person identification methods are using an ID card or the combination of a username and password. Recently, new techniques based on biometrics have been introduced so that people do not need to worry if they forget their username and password. For example, fingerprint and iris recognition are becoming common methods of person identification; however, users are usually required to interact with a system to use these traits. In some non-critical situations, it may be more convenient to utilize soft biometrics for person identification, although these features are not as unique for a specific person. In this work, we propose to conduct gait analysis that can be performed from a distance without disturbing user activities. We utilize depth cameras to capture user movements and create motion sequences. Then, a motion sequence is transformed to a motion string with appropriate data preprocessing and clustering techniques. Representative motion strings representing the individual behaviour of a user are retrieved and utilized to identify people. Empirical studies based on real motion data show that our approach performs well in person identification.

The representation of mythical characters in video games is the result of historical precedence as well as the designer's imagination. This study examines the relationship between classical allusions and the visual representation of video game characters, with a specific focus on the half-human figure of Medusa, to understand the mechanisms of character creation. Medusa is a character featured in Greek mythology; she is typically depicted as a winged woman with venomous snakes for hair. This study conducts an analysis of the representations of Medusa in literature, and then creates a novel visual representation based on this analysis as well as on gamer acceptance of character representations.

The histogram modification (HM) method proposed by Ni et al. is very efficient for reversible data hiding (RDH). Besides the excellent execution-time performance, Ni et al.’s HM-based RDH (HMRDH) method has a high PSNR lower bound of marked images. In this short communication, an observation on Ni et al.’s HM-based RDH (HMRDH) method is pointed out that the distortion of the marked image from Ni et al.’s method is dependent on the number of 1’s in the watermark. From this observation, we first present a watermark complement scheme to reduce the distortion occurred in Ni et al.’s HMRDH method. Later, combinatorial analysis for average distortion ratio of the proposed scheme is provided. This analysis motivates us to present a block-based complement scheme to improve the distortion reduction further. The tradeoff between the distortion and the number of partitioned blocks is also investigated. Taking nine well-known trademarks as the test watermarks and two cover images with different types of content, experimental results demonstrated the distortion reduction and higher PSNR lower bound advantages of the proposed block-based watermark complement scheme.

Advances in data acquisition, processing and visualization techniques have had a tremendous impact on medical imaging in recent years. However, the interpretation of medical images is still almost always performed by radiologists. Developments in artificial intelligence and image processing have shown the increasingly great potential of computer-aided diagnosis (CAD). Nevertheless, it has remained challenging to develop a general approach to process various commonly used types of medical images (e.g., X-ray, MRI, and ultrasound images). To facilitate diagnosis, we recommend the use of image segmentation to discover regions of interest (ROI) using self-organizing maps (SOM). We devise a two-stage SOM approach that can be used to precisely identify the dominant colors of a medical image and then segment it into several small regions. In addition, by appropriately conducting the recursive merging steps to merge smaller regions into larger ones, radiologists can usually identify one or more ROIs within a medical image.

This paper proposes the use of unified RGBD sensors in omni directional camera for constructing 3D maps of remote environments. The use of the unified sensor simplifies and accelerates the construction process. These advantages enable teleoperation systems to achieve faster refresh rate in visualizing 3D maps in operator's displays. Since the unified RGBD sensor technology is still unavailable for production, we validate the camera design on a simulation system. This work shows an alternative approach for the near future on constructing 3D maps, which is currently performed using sophisticated and extensive methods. We further experimented the camera design on a simulated teleoperation system, in which a user may control a mobile robot to navigate and to explore remote environments. The results showed that this camera is effective in providing spatial and color information for modeling remote environment into 3D maps.

Fractal image compression explores the self-similarity property of a natural image and utilizes the partitioned iterated function system (PIFS) to encode it. This technique is of great interest both in theory and application. However, it is time-consuming in the encoding process and such drawback renders it impractical for real time applications. The time is mainly spent on the search for the best-match block in a large domain pool. In this paper, a spatial correlation genetic algorithm (SC-GA) is proposed to speed up the encoder. There are two stages for the SC-GA method. The first stage makes use of spatial correlations in images for both the domain pool and the range pool to exploit local optima. The second stage is operated on the whole image to explore more adequate similarities if the local optima are not satisfied. With the aid of spatial correlation in images, the encoding time is 1.5 times faster than that of traditional genetic algorithm method, while the quality of the retrieved image is almost the same. Moreover, about half of the matched blocks come from the correlated space, so fewer bits are required to represent the fractal transform and therefore the compression ratio is also improved.

In this paper, a new mining capability, called mining of substitution rules, is explored. A substitution refers to the choice made by a customer to replace the purchase of some items with that of others. The mining of substitution rules in a transaction database, the same as that of association rules, will lead to very valuable knowledge in various aspects, including market prediction, user behaviour analysis and decision support. The process of mining substitution rules can be decomposed into two procedures. The first procedure is to identify concrete itemsets among a large number of frequent itemsets, where a concrete itemset is a frequent itemset whose items are statistically dependent. The second procedure is then on the substitution rule generation. In this paper, we first derive theoretical properties for the model of substitution rule mining and devise a technique on the induction of positive itemset supports to improve the efficiency of support counting for negative itemsets. Then, in light of these properties, the SRM (substitution rule mining) algorithm is designed and implemented to discover the substitution rules efficiently while attaining good statistical significance. Empirical studies are performed to evaluate the performance of the SRM algorithm proposed. It is shown that the SRM algorithm not only has very good execution efficiency but also produces substitution rules of very high quality.

Nowadays, recommender systems have become a necessity in various applications, especially in a large-scale online shop. In addition to the rating information provided by the users, social relationships of a user begin to be incorporated to further improve the performance of current recommender systems. Among several alternatives, matrix factorization is recognized as an effective technique to reduce data dimensionality and to capture significant latent relationships between users and items. Furthermore, recommender systems are used in an ever-changing commercial environment and usually operate on the large-scale data. Note that there are always new users, items and ratings as time advances, resulting in a rating matrix of increasing size. This poses a challenging problem because decomposing the entire matrix is costly. In this work, we thus propose an incremental scheme to directly update the rating matrix without the need to decompose the entire rating matrix. This helps to achieve better efficiency at the cost of some approximation errors. Experimental results show that our scheme has high efficiency as expected and significantly enhances the recommendation quality for cold-start users.

As low-cost depth cameras are released recently, motion data containing 3D coordinates of skeleton joints during a time period can be directly captured. Nevertheless, analyzing the motion data is usually a challenging problem and requires huge computation costs because of the high-dimensionality. Among several alternatives, the self-organizing map (SOM) is verified to be an effective technique to handle such motion data. Specifically, a captured motion sequence can be easily and precisely mapped to form an indexed motion string through the use of a trained SOM. However, the training process of the SOM is high computation complexity and is thus typically tedious. In view of this, we propose in this work to incorporate a hierarchical structure into the SOM technique. Generally speaking, our approach named as GQSOM (growing quadtree self-organizing map) helps to significantly reduce the required computation complexity while preserving the effectiveness of the SOM technique. Empirical studies using the WorkSU-10 exercise dataset show that our approach is both efficient and effective to perform indexing and retrieval tasks of motion data.

Developing intuitive interface to control humanoid robots with high degrees-of-freedom (DOF) in real-time remains as an essential issue. Since most humanoid robots have at least sixteen DOF, achieving real-time performance in controlling humanoid robot posture is hard to accomplish. This paper proposes a method to calculate joint angles of humanoid robots according to human's joint locations tracked by a motion capture system. The objective is to make the humanoid robot imitating human posture in real-time. Our proposed method requires a human motion capture system that could provide at least thirteen 3D joint positions: a pair of arm (shoulder, elbow, hand), a pair of leg (hip, knee, ankle), and a torso. Every joint that interconnects two limbs has to have two DOF. Our proposed method is implemented on RoboBuilder humanoid robot. The experimental results show that the robot can imitate human posture in real-time correctly. Moreover, the results verify that the proposed method is applicable to different type of humanoid robots using WeBots simulation software.

Advances in cell-fate mapping have revealed the complexity in phenotype, ontogeny and tissue distribution of the mammalian myeloid system. To capture this phenotypic diversity, we developed a 38-antibody panel for mass cytometry and used dimensionality reduction with machine learning-aided cluster analysis to build a composite of murine (mouse) myeloid cells in the steady state across lymphoid and nonlymphoid tissues. In addition to identifying all previously described myeloid populations, higher-order analysis allowed objective delineation of otherwise ambiguous subsets, including monocyte-macrophage intermediates and an array of granulocyte variants. Using mice that cannot sense granulocyte macrophage-colony stimulating factor GM-CSF (Csf2rb(-/-)), which have discrete alterations in myeloid development, we confirmed differences in barrier tissue dendritic cells, lung macrophages and eosinophils. The methodology further identified variations in the monocyte and innate lymphoid cell compartment that were unexpected, which confirmed that this approach is a powerful tool for unambiguous and unbiased characterization of the myeloid system.

By feeding personal e-mails into the training set, personalized content-based spam filters are believed to classify e-mails in higher accuracy. However, filters trained by both spam mails and personal mails may have difficulty classifying e-mails with the same characteristics of both spam and ham. In this paper, we propose a two-tier approach of using two filters trained only with either personal mails or spam mails. E-mails classified as legitimate mails by the legitimate mail filter may pass, while the remaining e-mails are processed by the spam filter in an ordinary way. Experiments in this paper are performed on two mail servers–one equipped with ordinary spam filter, and the other equipped both the legitimate mail filter and the spam filter. By combining the two filters with tuned thresholds, a much lower false positive rate is observed under the same false negative rate comparing to the ordinary filter.

We present a teleoperation system which supports foveated multi-resolution, wide field of view visual feedback. In the proposed system, a saliency-based visual attention system is implemented to simulate human visual system and to help determine the candidate visual attention area automatically. We link the visual attention system and the foveated multi-resolution image compression system to realize a real-time teleoperation system which requires less communication bandwidth to classic systems. Our experiment results show that the average operation time of the proposed system is only 57.9% comparing to a system with no visual attention clues in navigation tasks.

Multimodal image registration (MMIR) is the alignment of contents in images captured from different sensors or instruments. MMIR is important in medical applications as it enables the visualization of the complementary contents in biomedical microscopic images. The registration for such images can be challenging as the structures of their contents are usually only partially similar. Thus in this paper, we propose a new method to maximize the structural similarity of the contents in such images by utilizing intensity relationships among Red-Green-Blue color channels. Our experimental results will demonstrate that our proposed method substantially improves the accuracy of registering such images as compared to the state-of-the-art methods.

The purpose of this research is to develop multi-talented humanoid robots, based on technologies featuring high-computing and control abilities, to perform onstage. It has been a worldwide trend in the last decade to apply robot technologies in theatrical performance. The more robot performers resemble human beings, the easier it becomes for the emotions of audiences to be bonded with robotic performances. Although all kinds of robots can be theatrical performers based on programs, humanoid robots are more advantageous for playing a wider range of characters because of their resemblance to human beings. Thus, developing theatrical humanoid robots is becoming very important in the field of the robot theatre. However, theatrical humanoid robots need to possess the same versatile abilities as their human counterparts, instead of merely posing or performing motion demonstrations onstage, otherwise audiences will easily become bored. The four theatrical robots developed for this research have successfully performed in a public performance and participated in five programs. All of them were approved by most audiences.

As there are more and more surveillance cameras installed in public places, a challenging problem is to discover unusual behavior patterns from a huge amount of video data. However, this task is currently only feasible for human beings because both object recognition and intention detection are still difficult for computer vision. Recently, the development of low-cost depth cameras significantly improves the efficiency and effectiveness of capturing motion data. We thus propose in this work an algorithmic scheme that extracts unusual behavior patterns from motion capture data. Specifically, feature extraction and data clustering techniques are applied in our scheme so as to detect such outlier patterns. Example applications of our scheme include public area surveillance and home healthcare.

To ease the problem of data overloading, it is crucial to understand the user behavior when s/he interacts with online contents, or more specifically, a web page containing several entries for further exploration. We devise to estimate the relevance of an entry to the user goal by observing eye movements as implicit feedback. Specifically, this study proposes a framework that assumes eye movement measures can be used to infer a useri¦s cognition. A rating task was conducted in which subjects were required to judge whether an image was relevant to a word. Results showed that the total fixation duration and the fixation count can be used to discriminate between the relevant and irrelevant conditions, in contrast, the first fixation duration cannot. In addition, the subjective rating and relevancy manipulation interacted on the total fixation duration. Converging evidence verifies the assumption we have proposed.

With the advance of networking technologies, the widespread use of handheld devices not only enables instant access of required information but also shortens the distance among acquaintances. Nowadays one can easily share their photos on the Internet right after the photos are taken. With an increasing amount of images, a challenging problem for computer algorithms is to discover the visual-concept embedded within an image. Instead of modeling this problem as a classification process which attempts to categorize images into different pre-defined classes, we propose in this work to start with identifying similar image patches. Specifically, we adopt the technique of mining association rules to construct evident relationships among image patches so as to discover identical visual-concepts from online images.

As an inevitable trend, more and more robots are designed to be sold as household products in recent years. Famous examples like AIBO, RoboSapien, and Kondo, though aiming on different functionality respectively, are all affordable by general family. Among the robots stated above, four-legged robots have great advantage on locomotion over stair, uneven or multilevel floor, and floor with scattered stuff. Four-legged robots can also be used as a mechanical mule and are capable of carrying significant payloads, such as BigDog manufactured by Boston Dynamics (Raibert et al., 2008). In this chapter, we discuss the directional bias problem in depth and introduce an approach to dynamically detect the direction bias utilizing gait pattern information and the feedback of accelerator sensor. To evaluate how effective this approach is, experiments are performed on two Sony’s AIBO robots. There are lots of research topics of four-legged robots such as balance control, gait generation, image recognition, walking bias detection, to name but a few. In this chapter, we focus on bias detection technique of four-legged robots. Comparing to mechanical mule, AIBO robot is designed to be light weighted and is equipped with plastic hemisphere on its feet such that it does not scratch the walking plane such as beech solid wood floor. This design is a reasonable result to a household robot, but it also makes the robot not able to step firmly, thus produces directional bias even when walking straight on flat and smooth plane. Since the directional bias of some AIBO robots is obvious and this kind of bias tends to accumulates as long as the robot is walking, it would be nice if there is an algorithm to automatically detect and correct the walking directional bias in real-time. The most popular sensors used to detect heading direction of robots are video cameras, gyroscopes and accelerometers. Most image processing algorithms to detecting directional bias are time consuming and require more computing power than accelerometer based approaches. Since AIBO does not have gyroscope equipped, we choose to develop our algorithm according to accelerometer data. In theory, the distance of bias can be calculated by integrating the acceleration twice, but the acceleration data obtained from AIBO is not accurate enough to generate trustworthy data. Therefore, another data source is necessary to enhance our algorithm. After analyzed the characteristics of the three-axial acceleration sensors on AIBO, we confirmed the reliability of three-axial acceleration sensor on AIBO 17

Studies on color quantization have indicated that its applications range from the relaxation of displaying hardware constraints in early years to a modern usage of facilitating content-based image retrieval tasks. Among many alternatives, approaches based on neural network models are generally accepted to be very effective in color quantization. However, the inefficiency prevents their usefulness from practical usage. In view of this, we thus propose to incorporate a growing quadtree structure to the self-organizing map (GQSOM) technique in this work. Specifically, the strategy of inheriting from parent neurons hierarchically facilitates a much more efficient and flexible learning process. Both theoretical and empirical studies show that our approach is adaptive in determining an appropriate number of quantized colors. Moreover, the efficiency is significantly improved without compromise of the effectiveness.

The problem of image retrieval is much more challenging than that of text retrieval since significant features for representing and searching an arbitrary image vary with the image content. In this work, we focus on processing images containing landmarks which are easily recognizable by human users. Our observations show that through the use of proper techniques for skyline extraction, the retrieval task of landmark images can be feasible. Specifically, after conducting edge detection on an image, an approach which works in the similar fashion as depth-first search is proposed to extract the corresponding skyline precisely. In addition to theoretical analysis of relevant issues, empirical studies show that our approach is advantageous to be utilized in practical applications due to its high efficiency, effectiveness and scalability.

As the computer technology advances, data acquisition, processing and visualization techniques have a tremendous impact on medical imaging. On the other hand, however, the interpretation of medical images is still almost performed by radiologists nowadays. Developments in artificial intelligence and image processing show that computer-aided diagnosis emerges with increasingly high potential. In this paper, we develop an intelligent approach to perform image segmentation and thus to discover region of interest (ROI) for diagnosis purposes through the use of self-organizing map (SOM) techniques. Specifically, we propose a two-stage SOM approach which can precisely identify dominant color components and thus segment a medical image into several smaller pieces. In addition, with a proper merging step conducted iteratively, one or more ROIs in a medical image can usually be identified. Empirical studies show that our approach is effective at processing various types of medical images. Moreover, the feasibility of our approach is also evaluated by the illustration of image semantics.

In this paper, we propose an effective approach to detect the directional bias of walking four-legged robots. We found that light weight four-legged robots tend to have unstable directional bias even when walking on flat ground, thus this approach is designed to be able to dynamically and continuously detect the directional bias. By observing and analyzing data from acceleration sensors of AIBO robots, we found the relational model between the acceleration data and the walking gaits by which we may utilize to develop an ideal gait based approach to detect directional bias. In order to make sure the correctness of our approach, several experiments were designed and executed. As a result, we can correctly detect the direction of bias within five gaits, and the correct rate of single gait can be up to 90%.

We devise in this paper a regression-based algorithm, called algorithm FTP-DS (Frequent Temporal Patterns of Data Streams), to mine frequent temporal patterns for data streams. While providing a general framework of pattern frequency counting, algorithm FTP-DS has two major features, namely one data scan for online statistics collection and regression-based compact pattern representation.To attain the feature of one data scan, the data segmentation and the pattern growth scenarios are explored for the frequency counting purpose. Algorithm FTP-DS scans online transaction flows and generates candidate frequent patterns in real time. The second important feature of algorithm FTP-DS is on the regression-based compact pattern representation. Specifically, to meet the space constraint, we devise for pattern representation a compact ATF (standing for Accumulated Time and Frequency) form to aggregately comprise all the information required for regression analysis. In addition, we develop the techniques of the segmentation tuning and segment relaxation to enhance the functions of FTP-DS. With these features, algorithm FTP-DS is able to not only conduct mining with variable time intervals but also perform trend detection effectively. Synthetic data and a real dataset which contains net-Permission work alarm logs from a major telecommunication company are utilized to verify the feasibility of algorithm FTP-DS.

Studies on colour quantisation have indicated that its applications range from the relaxation of displaying hardware constraints in early years to a modern usage of facilitating content-based image retrieval tasks. Among many alternatives, approaches based on neural network models are generally accepted to be able to produce quality results in colour quantisation. However, these approaches using n quantised neurons require O(n) for a full search strategy, which is inefficient when n becomes large. In view of this, we propose to incorporate a growing quadtree structure into a self-organising map (GQSOM) which reaches a search time O(log n). Specifically, the strategy of inheriting from parent neurons hierarchically facilitates a much more efficient and flexible learning process. Both theoretical and empirical studies have shown that our approach is adaptive in determining an appropriate number of quantised colours, and the performance is significantly improved without compromise of the quantisation quality.