Three‐dimensional (3D) point cloud registration is a fundamental key issue in 3D reconstruction, 3D object recognition and augmented reality. In this study, the authors propose a novel local feature descriptor called local angle statistics histogram (LASH) for efficient 3D point cloud registration. LASH forms a description of local shape geometries by encoding their properties on angles between the normal vector of the point and the vector formed by the point and other points in its local neighbourhood. In addition, they propose a 3D point cloud registration algorithm based on LASH. The registration algorithm firstly detects triangle matching points with consistent similarity ratios, and then aggregates each pair of triangular matching points into a set of matching points. They can use these matching sets to calculate multiple transformations between two point clouds. Finally, they use the error function to identify the best transformation and to achieve coarse alignment of the two point clouds. Experiments and comparisons with other global algorithms demonstrate that the proposed approach can be applied to register point clouds with considerable or limited overlaps and is robust to noise.

This paper studies a set of MR technologies for middle school experimental teaching environments and develops a multi-channel MR user interface called Dream-Experiment. The goal of Dream-Experiment is to improve the traditional MR user interface, so that users can get a real, natural 3D interactive experience like real experiments, but without danger and pollution. In terms of visual presentation, we design multi-camera collaborative registration to realize robust 6-DoF MR interactive space, and also define a complete rendering pipeline to provide improved processing of virtual-real objects' occlusion including translucent devices. In the virtual-real interaction, we provide six interaction modes that support visual interaction, tangible interaction, virtual-real gestures with touching, voice, thermal feeling, and olfactory feeling. After users' testing, we find that Dream-Experiment has better interactive efficiency and user experience than traditional MR environments.

The most difficult problem of offline signature verification (SV) is that a signature is merely a static image missing a lot of the dynamic information associated with it. In this paper, three separate pseudo-dynamic features based on the gray level: gradient based local binary pattern (GLBP), statistical features of gray level co-occurrence matrix (SGLCM), simplified histogram of oriented gradients (SHOG) are proposed for writer-independent offline SV. These gray-level features can convey both texture information and the relative structural relationship of signature strokes. In addition, our experiments prove that the proposed features contain complementary information. Using random forests (RFs) as classifier, a fusion of the proposed features could achieve 7.42% and 0.08% average error rate (AER) for GPDS-253 and CEDAR datasets, respectively, which show the effectiveness of the proposed system. The implication of this paper is that part dynamic information could be extracted from a static gray level image.

CAPTCHA is a standard security technology to tell humans and computers and the most widely used method is text based scheme. As many text schemes have been broken, 3D CAPTCHAs have emerged as one of the latest one. In this paper, we study the robustness of 3D text-based CAPTCHA adopted by Ku6 which is a leading website providing videos in China and provide the first analysis of 3D hollow CAPTCHA. The security of this CAPTCHA scheme relies on a novel segmentation resistance mechanism, which combines Crowding Character Together (CCT) strategy and side surfaces which form the 3D visual effect of characters and lead to a promising usability even under strong overlapping between characters. However, by exploiting the unique features of the 3D characters in hollow font, i.e. parallel boundaries, the different stroke width of side faces and front faces and relationships between them, we propose a technique that segments connected characters apart and repairs some overlapped apart. The success segmentation rate is 70%. With minor changes, our attack program works well on its two variations, the segmentation rate is 75% and 85% respectively.

This paper studies offline text-independent writer identification of Chinese handwriting. The Bag of Features method is adopted for Chinese writer identification and performs much better than previous state-of-the-art methods. The feature adopted is scale invariant transform feature (SIFT) descriptor for it can extract local directional information from Chinese characters. Instead of Hard Voting, we use two newly devised coding strategies: Improved Fisher Kernels and Locality-constrained Linear Coding, to encode each SIFT descriptor. To make these coding strategies suitable to this new application area, absolute average pooling function is utilized. At last the K-nearest-neighbor classifier is used to identify the author of a handwriting image. Experimental results are conducted on a newly collected dataset of Chinese handwriting, CASIA Offline DB 2.1. Experimental results show our approach not only outperforms previous state-of-the-art methods, but also the traditional Bag of Word method using Hard Voting.

Background subtraction is a popular technique used in accurate foreground extraction with a stationary background. Since most outdoor surveillance videos are taken in complex environments, their "stationary" backgrounds change in some unknown patterns, which make the perfect foreground extraction very difficult. Based on visual background extractor (ViBe) scheme, in this paper we propose a new background subtraction algorithm which includes two innovative mechanisms and several other improved technique tricks. The paper inherits and develops background modeling based on pixel sample values, and use dynamic noise sampling and complementary learning to overcome the pixel-wise background model's intrinsic shortcomings. Besides, the algorithm works on the quantitative analysis without any estimation of the probability density function (pdf). Hence, it takes relatively low computational cost. Extensive experiments on a popular public dataset show that the proposed method has much better precision than ViBe, and could get the best precision and the highest average ranking compared with 27 state-of-the-art algorithms presented on the change detection website.

Handwritten Characters Recognition has long been a tough problem in pattern recognition and machine learning. Some special tasks, such as automatic check preprocessing, require Handwritten Chinese Legal Amounts recognition as a prerequisite. Since we expect to utilize machine instead of human to process bank checks, the recognition rate in such task must reach a relatively high rate. This paper proposes to use deep learning, auto-encoder as an effective approach for obtaining hierarchical representations of Isolated Handwritten Chinese Legal Amounts. Experiments show such representations are highly abstractive and can be used in character recognition. Meanwhile, a novel way by combining multiple Neural Networks in doing the work is proposed which proves to be able to improve the recognition rate significantly. This method reports a 0.64% error rate on a large test set over 10,000 samples and outperforms traditional methods using hand-crafted features and convolutional neural network committees (another deep learning model), narrowing the gap to human performance.

A hidden Markov model (HMM) based method for Chinese legal amount recognition is presented in this paper. In the training phase, gradient feature is extracted from sliding windows and character HMMs are trained with single character images. In the recognition phase, the text line image is segmented using sentence HMM, which is constructed by character HMMs according to a strict language model. The main difference between our proposed method and traditional methods is that our segmentation is guided by language model, which can solve many tough segmentation problems. Moreover, we combine the HMM-based method with traditional OCR method to improve the recognition accuracy. Experiments have been performed on 4,709 legal amount text line images extracted from real-life bank checks. The recognition rate is 97.13%.

We present an offline signature verification system using three different pseudo-dynamic features, two different classifier training approaches and two datasets. One of the most difficult problems of off-line signature verification is that the signature is just a static image while losing a lot of useful dynamic information. Three separate pseudo-dynamic features based on gray level: local binary pattern (LBP), gray level co-occurrence matrix (GLCM) and histogram of oriented gradients (HOG) are used. The classification is performed using writer-dependent Support Vector Machine (SVMs) classifiers and Global Real Adaboost method, where two different approaches to train the classifier. In the first mode, each SVM is trained with the feature vectors obtained from the reference signatures of the corresponding user and those random forgeries for each signer while the global Adaboost classifier is trained using genuine and random forgery signatures of signers that are excluded from the test set. The fusion of all features achieves the best result of 7.66% and 9.94% equal error rate in GPDS while 7.55% and 11.55% equal error rate in CSD respectively.

In this paper, we present a prototype system of courtesy amount recognition (CAR) for Chinese bank checks. The system deals with color check images and consists of five modules: skew correction, binarization, string extraction, segmentation and recognition. The whole system is designed under three principles: information fusion, complementary method combination and multi-hypotheses generation then evaluation. Information from color, gray and binary images is fused and complementary algorithms are applied for binarization and string extraction. Multi-hypotheses are made by keeping all possible candidates when ambiguous solutions exist in extraction and segmentation. Then the most suitable one is selected as the final result by evaluating the probabilities from recognition. Experiments show that our system is very promising based on a large number of real checks collected from different banks.

Because of the existence of possible carbon and seals, it's quite often that images of financial documents such as Chinese bank checks are suffered from bleed-through effects, which will affect the performance of automatic financial document processing such as seal verification and OCR. This paper presents an effective algorithm to deal with bleed-through effects existing in the images of financial documents. Double-sided images scanned simultaneously are used as in-puts, and the bleed-through effect is detected and removed after the registration of the recto and verso side images. There are two major aspects of contribution in our work. First, our algorithm can deal with images with complex background from real-life financial documents while most other algorithms only deal with images with simple background. Second, we combine the fast ICA algorithm with Gatos' local adaptive thresholding algorithm [1] to deal with the bleed-through effects. Experiments show that our proposed algorithm is very promising.

In this paper, we present a complete courtesy amount recognition system for Chinese bank checks. The system takes color bank check images as input and consists of three main processing steps: numeral string extraction, segmentation & recognition, and post-processing. They focus sequentially on: detection and extraction of numeral string; segmentation and recognition of the string; and further analysis of recognition results for acceptance or rejection. Information fusion, method complementarity, multi-hypotheses generation then evaluation are three principles employed for designing algorithms in the first two modules. And logistic regression is used for post-processing. A large number of real checks collected from different banks are used for testing the system. Read rate around 82% is observed when the substitution rate is set to 1%, which corresponds to that of a human operator. The performance can also be tuned further toward a suitable balance between inaccuracy and rejection, in accordance with user preference.

This paper presents a prototype system of Chinese payee name recognition (PNR) based on seal information in the back of Chinese bank checks. First, the seal imprints in the back images of Chinese bank checks are detected and extracted based on the color information. Second, the seal characters representing the payee name are segmented, rotated to horizontal position, and then recognized respectively. Third, the recognized seal characters are considered as the dictionary and payee name recognition is carried out as a verification process. Experiments demonstrate the effectiveness of our proposed method.

A novel binarization method for courtesy amount recognition of color Chinese checks is proposed in this paper. It includes two steps: background removal and seal removal. In the first step, two complementary binarization methods (Otsu method and Niblack method) are combined. In the second step, a multi-stage method for seal removal is proposed. All the seal-like pixels whose colors are similar to red in HSI color space are extracted first. Then the covered strokes which are extracted in the first stage are brought back to the left image by thresholding in RGB color space. Experiments show that the proposed method performs very well with little noise and can be used to remove background and seals for courtesy amount recognition of Chinese checks.

In this paper, we propose a new method on the segmentation and recognition of Chinese characters in both circular and elliptical seals for automatic Chinese seal imprint retrieval. The seal identification system compares an input seal image with its model seal. However, the model seal is usually found manually, which costs quite a lot of time and this manual stage has become the bottleneck of automatic seal identification. In addition, traditional automatic methods which were designed for circular seal imprint retrieval cannot work well for elliptical seal images. In order to overcome these shortcomings, our method first fits the contour of seal images and then automatically classifies them into either circular seal shape or elliptical seal shape. After that, we apply different methods for Chinese character segmentation and skew correction with circular and elliptical seal shape images respectively. Finally, we use our classifier combination strategy to recognize the Chinese characters on Chinese seal images. Once optical character recognition (OCR) is done, a model seal imprint corresponding to its input sample seal image is retrieved for verification from the seal imprint database. Experimental results show that our method has better performance than traditional ones.

A new method of logo detection in document images is proposed in this paper. It is based on the boundary extension of feature rectangles of which the definition is also given in this paper. This novel method takes advantage of a layout assumption that logos have background (white spaces) surrounding it in a document. Compared with other logo detection methods, this new method has the advantage that it is independent on logo shapes and very fast. After the logo candidates are detected, a simple decision tree is used to reduce the false positive from the logo candidate pool. We have tested our method on a public image database involving logos. Experiments show that our method is more precise and robust than the previous methods and is well qualified as an effective assistance in document retrieval.

A new method to extract crossing line features for off-line handwritten Chinese character recognition is proposed in this paper. Firstly, the input pattern is nonlinearly normalized in order to compensate for shape variations. Secondly, the normalized pattern is separated into four subpatterns according to the four kinds of elementary strokes. Thirdly, the four subpatterns are uniformly divided into M/spl times/M cells respectively. In every cell, the crossing lines are counted. Then a 4M/sup 2/-dimensional feature vector is generated. An off-line handwritten Chinese character recognition system is built based on this feature. Our experiments have demonstrated the effectiveness of the method proposed in this paper.