Yaokai Feng

Analyzing the Distribution of a Large-Scale Character Pattern Set Using Relative Neighborhood Graph

The goal of this research is to understand the true distribution of character patterns. Advances in computer technology for mass storage and digital processing have paved way to process a massive dataset for various pattern recognition problems. If we can represent and analyze the distribution of a large-scale character pattern set directly and understand its relationships deeply, it should be helpful for improving character recognizer. For this purpose, we propose a network analysis method to represent the distribution of patterns using a relative neighborhood graph and its clustered version. In this paper, the properties and validity of the proposed method are confirmed on 410,564 machine-printed digit patterns and 622,660 handwritten digit patterns which were manually ground-truthed and resized to 16 times 16 pixels. Our network analysis method represents the distribution of the patterns without any assumption, approximation or loss.

Analytical dynamic programming tracker

Visual tracking is formulated as an optimization problem of the position of a target object on video frames. This paper proposes a new tracking method based on dynamic programming (DP). Conventional DP-based tracking methods have utilized DP as an efficient breadth-first search algorithm. Thus, their computational complexity becomes prohibitive if the search breadth becomes large according to the increase of the number of parameters to be optimized. In contrast, the proposed method can avoid this problem by utilizing DP as an analytical solver rather than the conventional breadth-first search algorithm. In addition to experimental evaluations, it will be revealed that the proposed method has a close relation to the well-known KLT tracker.

SOM-based R*-tree for similarity retrieval

Feature-based similarity retrieval has become an important research issue in multimedia database systems. The features of multimedia data are useful for discriminating between multimedia objects (e.g., documents, images, video, music score, etc.). For example, images are represented by their color histograms, texture vectors, and shape descriptors. A feature vector is a vector that represents a set of features, and are usually high-dimensional data. The performance of conventional multidimensional data structures (e.g., R-tree family K-D-B tree, grid file, TV-tree) tends to deteriorate as the number of dimensions of feature vectors increases. The R*-tree is the most successful variant of the R-tree. We propose a SOM-based R*-tree as a new indexing method for high-dimensional feature vectors. The SOM-based R*-tree combines SOM and R*-tree to achieve search performance more scalable to high dimensionalities. Self-organizing maps (SOMs) provide mapping from high-dimensional feature vectors onto a two-dimensional space. The mapping preserves the topology of the feature vectors. The map is called a topological feature map, and preserves the mutual relationships (similarity) in the feature spaces of input data, clustering mutually similar feature vectors in neighboring nodes. We experimentally compare the retrieval time cost of a SOM-based R*-tree with that of an SOM and an R*-tree using color feature vectors extracted from 40,000 images.

Part-based methods for handwritten digit recognition

In this paper, we intensively study the behavior of three part-based methods for handwritten digit recognition. The principle of the proposed methods is to represent a handwritten digit image as a set of parts and recognize the image by aggregating the recognition results of individual parts. Since part-based methods do not rely on the global structure of a character, they are expected to be more robust against various deformations which may damage the global structure. The proposed three methods are based on the same principle but different in their details, for example, the way of aggregating the individual results. Thus, those methods have different performances. Experimental results show that even the simplest part-based method can achieve recognition rate as high as 98.42% while the improved one achieved 99.15%, which is comparable or even higher than some state-of-the-art method. This result is important because it reveals that characters can be recognized without their global structure. The results also show that the part-based method has robustness against deformations which usually appear in handwriting.

On the possibility of instance-based stroke recovery

This paper tackles the stroke recovery problem, which is a typical ill-posed reverse problem, by an instance-based method. The basic idea of the instance-based stroke recovery is to refer to the drawing order of a similar instance. The instance-based method has a strong merit that it can deal with multi-stroke characters and other complex characters without any special consideration. However, it requires a sufficient numbers of instances to cover those various characters. As an initial trial of the instance-based stroke recovery method, this paper describes the principle of the method and then provides several experimental results. The experimental results indicate the potential of the proposed method on recovering the drawing order of complex characters, as expected.

Character Image Patterns as Big Data

The ambitious goal of this research is to understand the real distribution of character patterns. Ideally, if we can collect all possible character patterns, we can totally understand how they are distributed in the image space. In addition, we also have the perfect character recognizer because we know the correct class for any character image. Of course, it is practically impossible to collect all those patterns - however, if we collect character patterns massively and analyze how the distribution changes according to the increase of patterns, we will be able to estimate the real distribution asymptotically. For this purpose, we use 822,714 manually ground-truthed 32×32 handwritten digit patterns in this paper. The distribution of those patterns are observed by nearest neighbor analysis and network analysis, both of which do not make any approximation (such as low-dimensional representation) and thus do not corrupt the details of the distribution.

Ag-Tree: a novel structure for range queries in data warehouse environments

In order to efficiently evaluate range-aggregate queries in data warehouse environments, several works on data cubes (such as the aggregate cubetree) are proposed. In the aggregate cubetree, each entry in every node stores the aggregate values of its corresponding subtree. Therefore, range-aggregate queries can be processed without visiting the child nodes whose parent nodes are fully included in the query range. However, the aggregate cubetree does not take range queries using partial dimensions and range queries without aggregation operations into account. That is, 1) a great deal of information that is irrelevant to the queries also has to be read from the disk for partially-dimensional range queries and 2) while it improves the performance of range queries with aggregate operations, it degrades the performance of the range queries without aggregate operations. In this paper, we proposed a novel index structure, called Aggregate-Tree (denoted as Ag-Tree), which gets rid of the above-mentioned weaknesses of the aggregate cubetree without any side effects. The experiments and discussions presented in this paper indicate that the new proposal is significant for range queries in data warehouse environments.

Efficient evaluation of partially-dimensional range queries using adaptive r*-tree

This paper is about how to efficiently evaluate partially-dimensional range queries, which are often used in many actual applications. If the existing multidimensional indices are employed to evaluate partially-dimensional range queries, then a great deal of information that is irrelevant to the queries also has to be read from disk. A modification of R*-tree is described in this paper to ameliorate such a situation. Discussions and experiments indicate that the proposed modification can clearly improve the performance of partially-dimensional range queries, especially for large datasets.

On the Possibility of Structure Learning-Based Scene Character Detector

In this paper, we propose a structure learning-based scene character detector which is inspired by the observation that characters have their own inherent structures compared with the background. Graphs are extracted from the thinned binary image to represent the topological line structures of scene contents. Then, a graph classifier, namely gBoost classifier, is trained with the intent to seek out the inherent structures of character and the counterparts of non-character. The experimental results show that the proposed detector achieves the remarkable classification performance with the accuracy of about 70%, which demonstrates the existence and separability of the inherent structures.

A Behavior-based Method for Detecting Distributed Scan Attacks in Darknets

The technologies used by attackers in the Internet environment are becoming more and more sophisticated. Of the many kinds of attacks, distributed scan attacks have become one of the most serious problems. In this study, we propose a novel method based on normal behavior modes of traffic to detect distributed scan attacks in darknet environments. In our proposed method, all the possible destination TCP and UDP ports are monitored, and when a port is attacked by a distributed scan, an alert is given. Moreover, the alert can have several levels reflecting the relative scale of the attack. To accelerate learning and updating the normal behavior modes and to realize rapid detection, an index is introduced, which is proved to be very efficient. The efficiency of our proposal is verified using real darknet traffic data. Although our proposal focuses on darknets, the idea can also be applied to ordinary networks.