Most of the existing ground-penetrating radar-based handheld landmine detectors use algorithms with low computational complexity to obtain detection results quickly. However, these algorithms do not show robust performance against background signals, such as noise signal, shake of sensor, and fluctuation of the signal. To ensure robustness against various background signals, a detection method using a model considering the statistical characteristics of signals should be used. Principal component analysis (PCA), which is a representative method, showed excellent performance in preprocessing and detection steps in the field of landmine detection. However, when the number of training data is small, the performance of PCA cannot be guaranteed. In addition, the amount of computation is increased because of high-dimensional data. Therefore, in this study, we propose a bilateral 2-D-discrete cosine transform PCA (B2D-DCT-PCA) method that shows robust performance with a small number of training samples and reduces computational complexity. In addition, landmine-background reconstruction error (LBRE) is introduced as an index for evaluating how well the difference between landmine and its background is distinguished to determine a suitable preprocessing method for the proposed method, and LBREs according to various preprocessing methods are compared. As a result, the frequency shifting showing the best performance is used. After optimizing the parameters of the proposed method, B2D-DCT-PCA, we confirm that it shows better performance compared with the conventional methods.

In the ground-penetrating radar (GPR) B-scan images, various noise sources are superimposed due to the ruggedness of the surface, sensor vibration, and multiple reflections. Additionally, the intensity of the received signals from small-size or low-metal-content landmines is low. Thus, it is difficult to accurately detect buried mines. In this letter, we propose an effective method to improve the B-scan image so that accurate landmine detection is possible even in such conditions. The proposed B-scan image enhancement method is comprised of two main techniques: an A-scan transformation based on frequency shifting and a background-landmine dissimilarity measurement using cumulative intensity distribution (CID). Based on frequency shifting, an A-scan transformation is devised to attenuate the strong ac component contained by the received A-scan signal. The CID-based dissimilarity is introduced to measure how an A-scan differs from a background model in the presence of a landmine. The proposed dissimilarity measure provides a robust response to a rugged ground surface and sensor vibration. For performance analysis, we compared our method with some conventional methods using the GPR data set acquired by an ultra wide band GPR sensor manufactured by Hanwha Systems Co., Ltd. We carried out various experiments and verified that the proposed method has a better performance than the conventional methods.

In this paper, we propose a novel approach for enhancing the contrast of back-light images. Although many contrast enhancement methods have been proposed in the literatures, they are still prone to fail especially in restoring the textural details of the under-exposed regions in back-light images. To cope with this problem, the proposed approach employs a new pixel-wise gain correction scheme that adopts a regional rank-1 constraint. Based on this scheme, the contrast of the back-light image is enhanced notably while allowing smooth intensity changes and suppressing intensity inversions around adjacent pixels. In the experiments, the effectiveness of the proposed method is demonstrated qualitatively and quantitatively comparing with the state-of-the-art methods.

In this letter, we present a novel DCT-based local binary descriptor for the dense matching of multiple view stereo (MVS). Recently, although much progress has been made in the field of MVS, a key component of which, i.e., dense matching, is still a challenging task because it has two difficult issues: robust matching over non-salient regions (e.g., lines and textureless regions) and fast matching of a large number of pixels. To deal with these issues effectively, in the proposed dense descriptor, 2D DCT-based local features are utilized to achieve high discriminative power even for the non-salient regions. A binary representation is adopted to increase the matching performance as well as accelerate the matching speed via the Hamming distance. In addition, the discriminability of binarized vectors is further improved by a space-frequency pooling scheme. Through extensive experiments on the benchmark datasets for MVS, we demonstrate the superiority of the proposed descriptor over the state-of-the-art descriptors in terms of accuracy and efficiency.

As image-based 3-D modeling is used in a variety of applications, accordingly, the compression of 3-D object geometry represented by multiple images becomes an important task. This paper presents a model-based approach to predict the geometric structure of an object using its visual hull. A visual hull is a geometric entity generated by shape-from-silhouette (SFS), and consequently it largely follows the overall shape of the object. The construction of a visual hull is computationally inexpensive and a visual hull can be encoded with relatively small amount of bits because it can be represented with 2-D silhouette images. Therefore, when it comes to the predictive compression of object's geometric data, the visual hull should be an effective predictor. In the proposed method, the geometric structure of an object is represented by a layered depth image (LDI), and a visual hull from the LDI data is computed via silhouette generation and SFS. The geometry of an object is predicted with the computed visual hull, and the visual hull data with its prediction errors are encoded. Simulation results show that the proposed predictive coding based on visual hull outperforms the previous image-based methods and the partial surface-based method.

In this paper, we focus on developing a novel visual target tracking system (TRACTOR) that enables reliable operation in realistic tracking scenarios. Although much progress has been made in the field of visual target tracking, there are still challenging scenarios in which even state-of-the-art trackers do not operate reliably. For instance, most trackers are prone to drift if a target moves abruptly in unexpected directions, or reappears after being fully occluded by the clutters or disappeared from the field of view of a camera. To cope with these scenarios effectively, the proposed tracking system subdivides the task of visual target tracking into two subtasks, i.e., tracking and detection. 1) For target-visible frames, the tracker builds a collaborative framework with the proposed appearance, observation, and motion models thereby achieving robust performance against unexpected motions and appearance changes of a target and 2) for target-invisible frames, the detector verifies continuously whether a target candidate is the lost target or clutter thereby reducing false target alarms effectively. In extensive experiments, the proposed tracking system shows very promising performance in comparison with state-of-the-art tracking methods.

In this paper, we propose a novel background subtraction method which enables reliable detection of foreground objects in a long surveillance video stream. Recently, although much progress has been made in the field of background subtraction, there are still challenging scenarios (e.g., high frequency motion of dynamic texture, non-stationary motion of camera, abrupt changes of illumination, etc.) in the long surveillance videos in which even state-of-the-art methods are often prone to fail. To cope with these challenging scenarios effectively, in the proposed method, a background model is initialized in a low-dimensional subspace and then updated periodically based on a novel recursive on-line (2D)2PCA algorithm developed in this paper. Moreover, a threshold map is also updated in a scene-adaptive manner for labeling each pixel in a scene either foreground or background independently. Based on this on-line framework, the background of a surveillance video stream is reconstructed over time, thereby facilitating the detection of foreground objects reliably. In extensive experiments, we demonstrate that the proposed background subtraction method can cope with the aforementioned challenging scenarios more favorably than the state-of-the-art methods.

In this paper, we present a novel color-to-gray image conversion method which preserves both color and texture discriminabilities effectively. Unlike previous approaches, the proposed method does not require any user-specific parameters for conversion. Moreover, the computational complexity is low enough to be applied to real-time applications. These breakthroughs are achieved by applying the ELSSP (Eigenvalue-weighted Linear Sum of Subspace Projections) method, which is proposed in this paper for the color-to-gray image conversion. Experimental results demonstrate that the proposed method is superior to the state-of-the-art methods in terms of both conversion speed and image quality.

In this paper, we propose a visual hull-based predictive framework for compressing depth data of a 3D object. Given a 3D object represented by layered depth image, we obtain a visual hull as a model. Then we compute and encode prediction residuals between the visual hull and the original 3D object. We first explain the merits of a visual hull in terms of prediction accuracy, computational costs, and efficient encoding of a model, and then we show the overall process of the proposed method.

Most video coding standards use the discrete cosine transform, known to be near optimal for original images, to transform prediction errors. Since the statistical characteristics of prediction errors are quite different from those of original images, a more suitable transform for prediction errors has to be devised. In this letter, we introduce a novel statistical model for inter-frame prediction error and propose an adaptive transform based on the model. In addition, in order to reduce the computation time, a fast and efficient algorithm is developed. Experiments on well-known image sequences confirm that our proposed transform can improve the performance of transform coding significantly.

A Layered Depth Image (LDI) is one of the popular representation and rendering methods for 3D objects with complex geometries. In this paper, we propose the new compression algorithm for depth information of a 3D object represented by LDI. For the purpose, we introduce the concept of partial surfaces to seek highly correlated depth data irrespective of their layer and derive a depth compression algorithm by using them. Partial surfaces are approximated by a Be¿zier patch and residual information is encoded by a shape-adaptive transform. Experimental results show that our proposed compression method achieves a better compression performance than any other previous methods.

To build a compact and photorealistic representation of a 3-D object without using polygonal mesh, 3-D representations of an object based on images accompanied by depth values were developed. In this letter, we propose a new compression system for the depth component of layered depth image (LDI), one of these representations, which represents a 3-D object. In order to achieve a high compression performance, it is necessary to group more correlated depth data. For this purpose we introduce the concept ofpartial surfacesfrom depth point cloud in LDI and derive a depth compression algorithm based on them generated from LDI. Neighboring depth points on the generated partial surface are likely to have high correlation, since it is a curved surface composed of surface points of the object. From experimental results, our proposed compression method outperforms other previous methods in terms of compression efficiency.

In this paper, we propose a computationally efficient space-varying dynamic range compression method based on statistical analysis. The proposed method reveals fine details without affecting already well-balanced region. For this goal, we introduce an effective segmentation method for distinguishing a damaged region from a well-balanced region. The statistical analysis and indirect use of the characteristics in segmented regions are introduced. In addition, to automate the full chain of dynamic range compression, we propose an enhancement condition which determines whether a certain image needs to be processed or not.

Frequency analysis of visual data is a fundamental step in understanding the characteristics of the data and developing necessary tools for processing it. When it comes to analyzing stereoscopic 3-D images, however, conventional methods for 2-D images need to be modified because stereoscopic 3-D images are created by a cognitive process of the human brain. We present the frequency domain analysis of a stereoscopic 3-D image with a geometry model of human binocular vision, and based on the analysis, we propose a disparity-adaptive anti-aliasing filter. From our analysis, we concluded that there is an inherent cause of aliasing due to frequency scaling by disparity. The analysis result is consistent with the knowledge obtained from past experiences, which supports the reliability of the analysis, and the proposed filter can be used for reducing undesirable visual discomfort.

Automatic white balance is a key function to remove undesirable color cast in digital cameras. Assumption-based methods (e.g. gray world method, white patch method) are widely used for the AWB process of digital cameras because of their low computational costs. However, they are highly dependent on the validity of their assumption. To solve this problem, we propose an adaptive feature extraction scheme for finding neutral color. The proposed method uses three features which are complementary and computationally effective. Moreover, we use the standard illuminants for evaluating an estimated illuminant not only to improve the accuracy of illuminant estimation, but also to provide the AWB skip mode. The experimental results show that the proposed method outperforms conventional assumption-based methods and arrives at reasonable color compensation over a wide range of scenes.

In this paper, a robust face recognition algorithm is proposed, which is based on the elastic graph matching (EGM) and discriminative feature analysis algorithm. We introduce a cost function for the EGM taking account of variations in face pose and facial expressions, and propose its optimization procedure. Our proposed cost function uses a set of Gabor-wavelet-based features, called robust jet, which are robust against the variations. The robust jet is defined in terms of discrete Fourier transform coefficients of Gabor coefficients. To cope with the difference between face poses of test face and reference faces, 2 x 2 warping matrix is incorporated in the proposed cost function. For the discriminative feature analysis, linear projection discriminant analysis and kernel-based projection discriminant analysis are introduced. These methods are motivated to solve the small-size problem of training samples. The basic idea of PDA is that a class is represented by a subspace spanned by some training samples of the class instead of using sample mean vector, that the distance from a pattern to a class is defined by using the error vector between the pattern and its projection to the subspace representing the class, and that an optimum feature selection rule is developed using the distance concept in a similar way as in the conventional linear discriminant analysis. In order to evaluate the performance of our face recognition algorithm, we carried out some experiments using the well-known FERET face database, and compared the performance with recently developed approaches. We observed that our algorithm outperformed the compared approaches.

Face detection in an image sequence is a challenging problem for many applications. In this paper, a novel face detection method is proposed. In order to detect faces in a sequence, based on Bayesian decision theory, we construct a unified framework of most face-like region selection, face/non-face classification, and detection result correction. And we propose face probability gradient ascent method to estimate the optimal position, scale, and rotation parameters of each face. In the experimental results, it is shown that the proposed method is more accurate and efficient than other conventional detection methods.

Change detection using a difference picture has been used to detect the location of moving targets and to track them. This method needs the assumption of a static camera, and global motion compensation is required in the case of a moving camera. We suggest a method for finding minimum bounding rectangles (MBRs) of moving targets in the image sequences using moving region detection, especially with a moving camera. If the global motion parameter is inaccurately estimated, the estimated locations of targets will not be accurate either. To alleviate this problem, we introduce the concept of the confidence measure and achieve more accurate estimation of global motion. With a moving camera, the new global motion estimation algorithm performs more precisely and it reduces the background compensation errors for change detection. Experimental results show that the proposed method successfully removes the background region and extracts MBRs of the targets.

We propose an efficient method to detect abrupt shot changes in a video sequence by using an unsupervised clustering. Most conventional shot change detection algorithms use only one kind of frame-by-frame difference feature such as pixel difference or histogram difference, so they can be applied to only specific situations. Another problem is the determination of appropriate threshold values to check the existence of shot changes. To overcome these problems we use several kinds of features simultaneously and propose a modified k-means clustering algorithm which changes the initial cluster center adaptively. Experimental results show that the proposed algorithm works well.

Bit allocation or the quantizer assignment problem is a basic and essential issue in lossy picture coding. The optimal solution for the bit allocation problem can be found by the Lagrangian method. However, it requires much computational time and memory. To reduce complexity overhead, we propose a fast scheme using a picture activity measure. Comparison among the existing activity measurement methods is presented to select the most reliable activity measure and the mapping relation between the activity value, and a quantization parameter is proposed.

The paper presents a morphological spatio-temporal segmentation method, which is based on a new similarity measure. This similarity measure considers jointly spatial and temporal information and consists therefore of two terms. The first term minimizes the displaced frame difference, considering the affine motion model. By the second term the spatial homogeneity of the luminance values of every region is maximized. The procedure toward complete segmentation consists of three steps: joint marker extraction, boundary decision, and motion-based region fusion. By incorporating spatial and temporal information simultaneously, we can obtain visually meaningful segmentation results. Simulation results demonstrates the efficiency of the proposed method.

This paper presents a new morphological spatio-temporal segmentation algorithm. The algorithm incorporates luminance and motion information simultaneously and uses morphological tools such as morphological filters and watershed algorithm. The procedure toward complete segmentation consists of three steps: joint marker extraction, boundary decision, and motion-based region fusion. First, the joint marker extraction identifies the presence of homogeneous regions in both motion and luminance, where a simple joint marker extraction technique is proposed. Second, the spatio-temporal boundaries are decided by the watershed algorithm. For this purpose, a new joint similarity measure is proposed. Finally, an elimination of redundant regions is done using motion-based region fusion. By incorporating spatial and temporal information simultaneously, we can obtain visually meaningful segmentation results. Simulation results demonstrates the efficiency of the proposed method.

In this paper, we propose a novel selective coding scheme based on layered structure. The scheme has 4 functional blocks that are target extraction, first layer coder for background region (FLC), second layer coder for target region (SLC) and buffer controller. For the target extraction process, we use motion vector grouping and change detection techniques, and increase the extraction accuracy of the target region by using a global motion compensation. The FLC and SLC are coders based on H.261 structure and generate low resolution background and high resolution target image, respectively. Simulation results show that the average PSNRs are obtained about 45.05 dB and 26.02 dB for target and background region respectively. Also the average number of bits generated in a frame is about 4780 bits: 2.16 bits and 41.84 bits per block are assigned to background and target region, respectively.

This paper presents a segmentation and motion estimation method for object-oriented analysis-synthesis coding. A major difficulty in estimating general motion is that it requires a large area of support in order to achieve a good estimation. Unfortunately, when the supporting area is large it is very likely to have multiple moving objects. To solve this problem, we propose a multi-stage segmentation method which is based on optical flow. The basic concept is to group homogeneous subregions with respect to simpler mapping model into large homogeneous regions with respect to more complex mapping model. By applying a hierarchy of mapping parameter model progressively, we can segment the whole changed region into several parabolic patches. Especially person's face in head-and-shoulder images can be described as one object.

This paper proposes a 3-D filtering scheme which reduces temporal aliasing effectively and requires much less memory than conventional schemes in hardware implementation. Our 3-D filter is implemented in the form of a space-variant 2-D finite-duration impulse response (FIR) filter in the spatial domain with motion information. The filter requires only one frame and its motion vectors. The simulation result for the image sequence with constant motion informs explicitly that our filter has temporal de-aliasing property and 3-D band-limitation.>