Global three-dimensional wind measurement is important for weather prediction' climate studies' and various meteorological studies. Current space-based passive sensors can make wind measurements with a large coverage area and high temporal and horizontal resolutions but has a low vertical resolution. The World Meteorological Organization (WMO) wants to develop space-based wind profiling systems. A space-based Doppler Wind Lidar (DWL) is one of useful and promising remote sensing techniques to fill the gap of current observations. The National Institute of Information and Communications Technology (NICT) is developing a single-frequency high-energy Tm,Ho:YLF laser, 1-μm key technology and instrument for a future space-based coherent DWL. In the paper' we will describe recent progress and results of development of 1-μm key technology.

In this paper, we propose a novel method to train deep neural networks for small-scale training dataset. Since we focus on the situation in which the training dataset is limited, increasing the number of training data by data augmentation is the most straight-forward solution. Instead of augmenting data at an input layer as in typical data augmentation, we adversarially augment features at hidden layer by adding small perturbations to the original features extracted from training data. We call the augmented features as adversarial features. To effectively avoid overfitting of the trained network, the perturbation is designed to be adversarial, which means that they are designed to significantly change the output of the network. Moreover, to induce the adversarial feature to be more reasonable as real data, we adopt a coefficient layer to constrain the adversarial feature to be represented by linear combination of the original features. Experimental results on several benchmark datasets show that our method substantially improves the performance of the deep neural network.

We consider the problem of generative pedestrian modeling for capturing common behaviors constituting trajectory dataset. We present a model, that simultaneously represents the trajectory data and latent dynamics associated with different behaviors. Model represents the trajectory dynamics as scaled component of cluster dynamics, and overall the cluster dynamics is shared among all trajectories belonging to a cluster, thus giving rise to similarity. Cluster dynamics is modeled by incorporating Bayesian nonparametrics, particularly the usage of Dirichlet process mixture model approach, which relaxes the number of unique behaviors or clusters. Additionally, the relative velocity scaling term encapsulates the relative nature of an individual trajectory to its cluster dynamics. Model parameters and latent states are inferred using sequential blocked Gibbs sampler, which can be scaled to large datasets.

A novel fast k-nearest neighbor (k-NN) search method is proposed for the face identification task. It is well suited for this task because (1) it works well with high dimensionality, (2) it can be used with various similarity scores such as inner product, Euclidean distance, and correlation coefficient, (3) it can achieve not only fast exact k-NN search but much faster approximate search, and (4) it does not require any training or special data structure, resulting in low maintenance cost for the target database. Similarity scores between query and target samples are aggregated sequentially along with their dimensions, and target samples with no possibility of being included in k-NNs are rejected. The possibility is evaluated on the basis of the upper and lower bounds of the score for residual dimensions. Experimental results for a face database demonstrated that the proposed method achieves equal or better accuracy than other methods and is ten times faster than an exhaustive search with no degradation in the rank-k identification rate.

In this paper, we propose a novel scheme for domain adaptation in which feature transform and instance weights are jointly optimized. Due to the joint optimization, we can obtain feasible feature transform for domain adaptation while we jointly eliminate source samples which are unrelated to target samples by estimating those weights. By introducing regularization which induces the weights to be homogeneous, we can increase the number of successfully adapted source samples as much as possible resulting in the stable training of classifiers after domain adaptation. Experimental results on both benchmark data and real surveillance video show that our method can achieve the same or better performance than that of state-of-the-art methods, though we used only the simplest feature transform, that is linear transform, in our method.

This paper presents a novel image-reconstruction (IR) method for satellite imagery. To reduce unnatural artifacts, which often appear in existing IR methods, we propose a novel scheme to accurately estimate PSF and a novel regularization for IR. In the PSF estimation, hyper-parameters are estimated to reduce artifacts while improving sharpness based on our new criterion, which employs a residual of sigmoidal-function fitting to a strong edge on the reconstructed image as measurement of the amount of the artifacts. After the PSF estimation, we conduct IR based on the estimated PSF. In IR process, we employ a novel regularization that induces gradients of the reconstructed image to be close to those of its guide image. Since the guide image contains only the dominant structure of the input image without artifacts, the regularization leads to fewer artifacts. In addition, we employ a learning-based method for setting the spatially adaptive strength of the regularization effectively. Experimental results on real satellite imageries show that our method works better than other state-of-the-art IR methods.

In this paper, we propose a feature selection method that uses graph cuts based on both relevance and redundancy of features. The feature subset is derived by an optimization using a novel criterion which consists of two terms: relevance and redundancy. This kind of criterion has been proposed elsewhere, but previously proposed criteria are hard to optimize. In contrast, our criterion is designed to satisfy submodularity so that we can obtain a globally optimal feature subset in polynomial time using graph cuts. Experimental results show that the proposed method works well, especially in the case of a medium-size subset where existing approaches are weak because of the many possible feature combinations.

This paper presents a new learning formulation for classifier design called ``General Loss Minimization.'' The formulation is based on Bayes decision theory which can handle various losses as well as prior probabilities. A learning method for RBF kernel classifiers is derived based on the formulation. Experimental results reveal that the classification accuracy by the proposed method is almost the same as or better than Support Vector Machine (SVM), while the number of obtained reference vectors by the proposed method is much less than that of support vectors by SVM.

This paper describes a method for reducing the number of candidates in Kanji character recognition. Such reduction is vital to increasing the speed of such applications as Kanji address recognition. Further, it also reduces the probability of misreadings in linguistic postprocessing. First, we define a confidence value by which we can express the potential correctness of recognition candidates. Next, we define "accumulated confidence value" and use it as a measure to express a threshold for candidate acceptance in Kanji character recognition. The efficiency of the proposed method is evaluated in an experiment using IPTP CD-ROM2 Japanese address images. Results show that a roughly 31% reduction in the number of candidates was obtained without reducing the number of correct candidates.

This paper presents a novel learning method for definite canonicalization (DC) based on minimum classification error (MCE). It is shown that DC is identical to normalized cross-correlation, and that the complementary similarity measure is derived from DC for binary patterns. The proposed learning method is derived from the framework of generalized learning vector quantization (GLVQ), which is one of the discriminative learning methods based on MCE. Experimental results obtained for machine-printed Kanji character recognition reveal that the proposed method achieves high performance recognition of low-quality patterns.

This paper reports a formulation of learning vector quantization (LVQ) using a new misclassification measure based on minimum classification error (MCE). We show that the convergence property of reference vectors depends on the definition of the misclassification measure, and show that our definition guarantees the convergence, unlike LVQ1.1 or Juan and Katagiri's formulation based on MCE (1992). Experimental results for handwritten digit recognition reveal that the proposed method is superior to LVQ algorithms in recognition capability.

This paper presents a realtime pedestrian counting system based on x-t spacetime image analysis. The system counts the number of pedestrians through the combination of three processes: moving object extraction, object counting by region labeling, and direction detection by flow estimation. In order to extract objects reliably under various illumination changes, we propose new moving objects extraction methods that extract moving regions whose shadows have been eliminated by using adaptive background image reconstruction and the color information processing of the images. For labeling and flow estimation, we use the Ortho-sectioning method that analyzes extracted regions on a spacetime slice image of the original three dimensional volume. Experimental results confirm the robustness of the system against illumination changes.