Despite extensive researches for face recognition (FR), it is still difficult to apply deep CNN models to NIR FR due to a lack of training data. In this study, we propose a fine-tuning approach to allow deep CNN models to be applied to NIR FR with small training datasets. In the proposed approach, parameters of deep CNN models for RGB FR are utilized as initial parameters to train deep CNN models for NIR FR. The proposed approach has two main advantages: 1) High NIR FR performances can be achieved with very small public training datasets. 2) We can easily secure good generalization for NIR FR in various environments. Our fine-tuning approach achieved a validation rate of 99.70% with the PolyU-NIRFD database. In addition, we constructed private face databases with Intel® RealSense™ SR300. On the VF_NIR database, which is one of the private databases, we achieved a validation rate of 94.47%.

Video synopsis allows us to analyze security videos efficiently by condensing or shortening a long video into a short one. To generate a condensed video, moving objects (a.k.a. object tubes) in the video are rearranged in the temporal domain using a predefined objective function. The objective function consists of several energy terms which play important roles in making a visually appealing condensed video. One of the energy terms, collision energy, creates a bottleneck in the computation because it requires two object tubes to calculate the degree of collision between them. Existing approaches try to reduce the computation time of the collision energy calculation by reducing the number of tubes processed at once. However, those approaches are not sufficient to generate condensed video when the number of object tubes becomes large. In this letter, we propose a fast Fourier transform (FFT)-based parallelized tube rearrangement algorithm. To take advantage of both parallel processing and FFT, we represent object tubes as three-dimensional binary matrices (occupation matrices). An objective function of the tube rearrangement problem is defined on the occupation matrix, and a starting position for each tube in the temporal domain is then determined by optimizing the objective function. Throughout the experiments, the proposed algorithm took a much shorter time to condense the video than existing algorithms, while other performance metrics were similar.

Local feature descriptors based on gradient orientation histogram show good performance even when images contain distortions such as view point change, blur and rotation. However their performance declines significantly when images are compressed using the block DCT based algorithm. Since images and videos are usually encoded to a compressed file format to reduce file size, many image processing applications inevitably treat compressed images. In this paper, we investigate the robustness of Zernike moment against compression noise. In our experiment using the INRIA dataset, we compared the matching results of the descriptors using the magnitudes of Zernike moments with SIFT descriptor in terms of recall vs. 1-precision metric. Magnitudes of Zernike moments provided better matching performance than SIFT when images contain compression noise.

In this paper we propose the method that extracts the semantic keyword from digital images automatically using color and texture features. The image semantic keyword is widely used in research area like image retrieval, categorization, annotation, management. The method consists of two steps: feature extraction and classification module. In order to extract feature, the image color and PACT (Principal component analysis of Census Transform) histogram are used. For classification, SVM (Support Vector Machine) classifier is used. The final keyword is annotated after post-processing. Experimental results indicate that the proposed method can accurately extract the image semantic keywords.

With popularization of multimedia devices, semantic analysis of sports video has been widely studied. In this paper, we propose a highlight generation method for basketball games. To create a video highlight, the proposed method selects interesting shots by modeling excitements of the game using score information. For this purpose, a video is first segmented into shots and classified as play and nonplay shots. At the same time, score of the game is automatically extracted from video frames. To select interesting shots, which should be included to the highlight video, excitement of the game is estimated from the variation of game scores. Unlike previous event-based video analysis methods which focus on individual events, our proposed method is able to reflect the contents of a game by considering excitements. Also, because the excitement modeling uses score information only, it can easily be extended to other types of sports video which have scores.

An automated vision system that accurately measures the thickness of a brake shoe for rolling stock is proposed. Images of the brake shoes are captured automatically while rolling stock is passing by an inspection station. Toward this goal, location of the brake shoe in the image has to be determined first because they are not the same in the captured images. Shadow regions between the brake shoes and wheels are utilized that are common in all captured images. The boundary of the shadow regions is modeled by an inverse polynomial, and the constrained curve fitting method is adopted. Then, two curves that correspond to the front and back of the brake shoe are determined using the initial curve and intensity change information. Finally, the thickness is calculated using the detected curves, and experimental results showed that the brake shoe thickness was measured with an accuracy of 0.654 mm.

Angular radial transform (ART), which is the region-based shape descriptor of MPEG-7, has desirable properties for representing shape information in a small number of features with no redundancy. However, in order for ART to be useful, especially in limited computing environments, the computational cost of ART must be greatly reduced. In this paper, we derive symmetric/antisymmetric properties from the basis functions of ART and present a fast and efficient method to compute the ART coefficients using these properties. The proposed method significantly reduces the number of sinusoidal operations and multiplications in computing the coefficients of ART. Moreover, the memory requirements needed to store the ART basis functions in lookup tables are only 25% of the conventional method. The experimental results are presented to show the effectiveness of the proposed method.

The conventional method of estimating the rotation angle of a pattern using the principal axes is not suitable for circular symmetric patterns since their eigenvalues are similar in both directions. In the paper, a robust method of estimating a rotation angle using the phase information of Zernike moments is presented. The experimental results show that the proposed method estimates the rotation angle of the circular symmetric patterns more accurately than the principal axes method, even in the presence of noise.

In this paper, we present a new reversible decorrelation method of three-dimensional (3-D) medical images for progressive transmission. Progressive transmission of an image permits gradual improvement of image quality while being displayed. When the amount of image data is very large, as a 3-D medical image, the progressive transmission plays an important role in viewing or browsing the image. The data structure presented in this paper takes account of interframe correlation as well as intraframe correlation of the 3-D image. This type of data structure has been termed the 3-D hierarchy embedded differential image (3-D-HEDI) as was derived from the earlier HEDI structure. Experiments were conducted to verify the performance of 3-D HEDI in terms of the decorrelation efficiency as well as the progressive transmission efficiency. It is compared with those of conventional hierarchy interpolation (HINT), two-dimensional (2-D) HEDI and differential pulse code modulation (DPCM). Experimental results indicate that 3-D HEDI outperforms HINT, 2-D HEDI and DPCM in both decorrelation efficiency as well as the progressive transmission efficiency on 3-D medical images.

An ever increasing number of registered trademarks has created greater demand for an automatic trademark retrieval system. We present a method for such a system based on the image content using shape features. Zernike or pseudo-Zernike moments of the image are employed as a feature set. To retrieve similar shapes, we take into account visually salient features that dominantly affect the global shape of the trademarks and ignore their minor detail. Experimental results on a database of 3,000 trademark images demonstrate that the proposed method retrieves visually similar trademarks which agree well with human perception.

The purpose of this paper is to present a new data structure for viewing large size medical images. The data structure is specifically designed suitable for lossless compression and viewing images interactively even when the channel bandwidth is limited as low as those of modems. Our data structure has been tested on several 2 K/spl times/2 K images by using various baud rates from 9600 to 28800. The thrust of this paper is two fold. First, browsing large image data files through limited bandwidth channels requires progressive transmission. Using conventional data structures, 1/3 more memory space is required to build the hierarchical structure. Our data structure utilizing two-dimensional DPCM embedded with hierarchy does not require extra memory space to build the structure for progressive transmission. The second important aspect of the data structure is the local zooming capability of the data for interactive viewing. This is achieved by rearranging the transmission order of image data so that the data in the region of interest can be transmitted first with the rest of the information on hold to be sent later.

Algorithms for constructing differential images with hierarchical data structure are presented. The data structures are simple, efficient, and ideal for viewing images in progressive transmission using lossless compression. Unlike conventional pyramidal structures, the total number of nodes required to build the structure is the same as the number of pixels in an image at the same time its hierarchy is preserved. These structures are constructed using subsampling or mean-sampling methods for predictors with block sizes of 2/spl times/2 or 3/spl times/3. Experiments were conducted to compare these structures in terms of their first order entropy and RMS errors in the reconstruction process. Results indicate that the mean-sampling with circular-difference method yields the lowest entropy, comparable to that with 1-D lossless DPCM predictive coding. Lastly, hardware for the efficient construction and access of the hierarchical structures is discussed and evaluated.>

We present a practical pattern recognition system that is invariant with respect to translation, scale and rotation of objects. The system is also insensitive to large variations of the threshold used. As feature vectors, Zernike moments are used and we compare them with Hu's seven moment invariants. For a practical machine vision system, three key issues are discussed: pattern normalization, fast computation of Zernike moments, and classification using k-NN rule. As testing results, the system recognizes a set of 62 alphanumeric machine-printed characters with different sizes, at arbitrary orientations, and with different thresholds where the size of the characters varies from 10/spl times/10 to 512/spl times/512 pixels.>

The authors show how large efficiencies can be achieved in model-based 3-D vision by combining the notions of discrete relaxation and bipartite matching. The computational approach presented is capable of pruning large segments of search space-an indispensable step when the number of objects in the model library is large and when recognition of complex objects with a large number of surfaces is called for. Bipartite matching is used for quick wholesale rejection of inapplicable models and for the determination of compatibility of a scene surface with a potential model surface taking into account relational considerations. The time complexity function associated with those aspects of the procedure that are implemented via bipartite matching is provided. The algorithms do not take more than a couple of iterations, even for objects with more than 30 surfaces.>