Many recent text recognition studies achieved successful performance by applying a sequential-label prediction framework such as connectionist temporal classification. Meanwhile, regularization is known to be essential to avoid overfitting when training deep neural networks. Regularization techniques that allow for semi-supervised learning have a greater impact than those that do not. Among widely researched single-label regularization techniques, virtual adversarial training (VAT) performs successfully by smoothing posterior distributions around training data points. However, VAT is almost solely applied to single-label prediction tasks, not to sequential-label prediction tasks. This is because the number of possible candidates in the label sequence exponentially increases with the sequence length, making it impractical to calculate posterior distributions and the divergence between them. Investigating this problem, we have found that there is an easily computable upper bound for divergence. Here, we propose fast distributional smoothing (FDS) as a method for drastically reducing computational costs by minimizing this upper bound. FDS allows regularization at practical computational costs in both supervised and semi-supervised learning. An experiment under simple settings confirmed that upper-bound minimization decreases divergence. Experiments also show that FDS improves scene text recognition performance and enhances state-of-the-art regularization performance. Furthermore, experiments show that FDS enables efficient semi-supervised learning in sequential-label prediction tasks and that it outperforms a conventional semi-supervised method.

We propose a navigation interface, which enables users to control locomotion and the speed intuitively in virtual environments constructed with spherical images. In these virtual environments, users can only pass through lines where the camera has passed through during recording. Therefore, it is necessary to inform users of the accessible directions. Furthermore, velocity control is needed for close exploration in virtual environments. Therefore, we propose the Motive Compass input interface, which intuitively shows users the accessible directions and enables them to control the velocity of locomotion. We conducted a large-scale demonstration experiment in a real exhibition and showed that our interface more effectively presents accessible directions and enables users to control their velocities.

In order to apply to land mine detection effectively, bistatic GPR using an optical electric field sensor as a receiver has been developed. The optical electric field sensor is very small and uses optical fiber instead of metallic coaxial cable. With the combination of these advantages and the bistatic radar system, it can be possible for an operator to measure quite flexible and safely. The sensor has been tested in stepped frequency radar system with frequency which consists of a vector network analyzer, a fixed double ridged horn antenna as transmitter and computer controlled receiver-positioner for 2D scanning. For considering effectiveness in real field, we applied impulse radar system, which consist of a digital oscilloscope and an impulse generator to produce the impulse with 250 psec pulse width and the diffraction stacking has been adopted for 3D image reconstruction. Detection of a PMN2 mine model was carried out by the impulse radar system at a sand pit with two modes of data acquisition: the stepped data acquisition and the continuous data acquisition. The PMN2 were detected clearly with sufficiently high resolution in TM mode measurement in both modes, the target contrast was almost the same while the scanning time decreased down to 1/18, 9 min/m2 in case of continuous data acquisition mode which satisfies the requirement for practical use.

In this paper, we propose a new group demodulator based on a complex multirate filter bank for multiple-channel mobile communication systems, which can handle asynchronously frequency-multiplexed signals. In addition, we propose a fast channel searching method to find active channels suited for the group demodulator. Computer simulation results show that the proposed group demodulator can achieve a relatively low hardware complexity keeping a good bit error rate performance.