Dynamic background estimation (DBE) and background subtraction (DBS) is a very effective way for moving object tracking when (a part of) background is continually fluctuating. In DBE, newest time-series video frames are kept and the pixel values at the same position in those frames are used to determine the value of background pixels. We propose an accurate and fast DBE/DBS method by defining a chromaticity-based color space which is less affected by brightness change, and by defining a similarity metric based on the distance in the color space. In our DBE, the pixel with the highest average similarity is determined as the background. By utilizing the statistical value derived from the calculation of the average similarities, the threshold value for our DBS after our DBE is adaptively determined. In the evaluation of Precision-Recall relation and F-measure of foreground extraction with public data set, our method with a median filter as pre-process showed comparable or better performance than other representative methods. The processing speed for VGA images was 24 fps that is much faster than other methods. Our method with resizing before a median filter showed about 10 point higher F-measure than others and 100 fps of the speed so that it is applicable to real-time object tracing systems.

View full abstract

This study aims to develop an estimation method for a 'shape representation map'. In this study, shape representation map is a nonlinear subspace formed by a set of shapes, in which the continuous change in shapes is naturally represented. By estimating the shape representing map, it is expected that the intrinsic parameter causing the shape change appears. In this paper, we propose a method using the power of SOM (SOMⁿ) to estimate the shape map. We carried out two simulations with the proposed method. One is the classification task of contours which are deformed by affine transformation, and the other is the localization task of a mobile robot by the skyline shape. The results showed that this method is useful when one needs to get information from shape changes.

View full abstract

Recently, the applications of rough set theory to real problems have been studied actively. However, when the number of induced decision rules by rough set approach is large, finding useful decision rules becomes often a formidable task. This fact may become a bottlenek in actual use of induced decision rules. In this study, we develop the decision rule visualization system for supporting the discovery of useful decision rules. The system visualizes decision rules having same conclusions in three-dimensional space using co-occurrence rates between atomic formulae and conclusions, those between atomic formulae and Hayashi's quantification method IV. An evaluation experiment is conducted. The results show that useful decision rules are discovered efficiently from many decision rules induced by the rough set approach. Some additional functions are implemented to the system in order to overcome the problems pointed out by examinees.

View full abstract

In this paper, we propose a system which produces funny proverbs. This proposed system uses the punch line framework named Sukashi. That is, by changing the end of the line in proverbs defying the prediction of users, the proposed system produces a funny sentence. The proposed system has 3 distinct features: First, the novel framework named Sukashi is employed. Second, various and huge number of Sukashi candidates are generated due to utilizing the Japanese Google N-grams being constructed from 255 billion words. Third, the factors of laughing -imageability and concrete level of the targeted words- are newly added to the sound length, difference of sound and concepts, which are frequently used in pun generation. Due to these factors, funnier Sukashis are able to be selected from candidates using fuzzy rules. We confirmed that the proposed system is able to generate Sukashis quite comparable to human-made ones in terms of both funny level and unpredictable level.

View full abstract

Interactive EC Fitting is an effective method to set the signal processing parameters of a hearing aid optimal to the preference of a user. However, it has problems that uncomfotable sounds are sometimes generated because of the fluctuation of user's evaluation and the evolutionary operations of EC and that the fitting takes a good amount of time due to the requirement of iterative user's evaluation. We thus propose an improved method that automatically evaluates candidate settings using the preference of sound volume apriori given by the user, and presents the user only candidate settings predicted as comfortable. We developed the improved method and conducted simulation-based experiments for examining the optimal number of candidate settings with no user evaluation and the effectiveness of the improved method. Under the condition that the number of candidate settings with user evaluation was fixed as 20, which is generally used in Interactive EC Fitting, the best performance was achieved for the number of candidate settings with no user evaluation of 10 in our experiments. With this number, the improved method significantly reduced the variance of evaluation values, and it is suggested that outliers, i.e., uncomfortable settings, were efficiently reduced. In addition, the improved method significantly accelerated the EC convergence, and the shortening of fitting time was suggested. Our future work will be some practical experiments using human subjects to examine whether the improved method reduce the users' fatigue.

View full abstract

In this paper, we propose Chain Form Reinforcement Learning for a reinforcement learning agent that has small memory. In the real world, learning is difficult because there are an infinite number of states and actions that need a large number of stored memories and learning times. To solve a problem, estimated values are categorized as “GOOD” or “NO GOOD” in the reinforcement learning process. Additionally, the alignment sequence of estimated values is changed as they are regarded as an important sequence themselves. We conducted some simulations and observed the influence of our methods. Several simulation results show no bad influence on learning speed.

View full abstract