Noboru Sugie

Linear-time connected-component labeling based on sequential local operations

This paper presents a fast algorithm for labeling connected components in binary images based on sequential local operations. A one-dimensional table, which memorizes label equivalences, is used for uniting equivalent labels successively during the operations in forward and backward raster directions. The proposed algorithm has a desirable characteristic: the execution time is directly proportional to the number of pixels in connected components in an image. By comparative evaluations, it has been shown that the efficiency of the proposed algorithm is superior to those of the conventional algorithms.

Neural edge enhancer for supervised edge enhancement from noisy images

We propose a new edge enhancer based on a modified multilayer neural network, which is called a neural edge enhancer (NEE), for enhancing the desired edges clearly from noisy images. The NEE is a supervised edge enhancer: Through training with a set of input noisy images and teaching edges, the NEE acquires the function of a desired edge enhancer. The input images are synthesized from noiseless images by addition of noise. The teaching edges are made from the noiseless images by performing the desired edge enhancer. To investigate the performance, we carried out experiments to enhance edges from noisy artificial and natural images. By comparison with conventional edge enhancers, the following was demonstrated: The NEE was robust against noise, was able to enhance continuous edges from noisy images, and was superior to the conventional edge enhancers in similarity to the desired edges. To gain insight into the nonlinear kernel of the NEE, we performed analyses on the trained NEE. The results suggested that the trained NEE acquired directional gradient operators with smoothing. Furthermore, we propose a method for edge localization for the NEE. We compared the NEE, together with the proposed edge localization method, with a leading edge detector. The NEE was proven to be useful for enhancing edges from noisy images.

A model-based sound localization system and its application to robot navigation

This paper describes a mobile robot equipped with a real time sound localization system as well as a sonar system for obstacle detection. The sound localization method is based on a model of the precedence effect of the human auditory system to cope with echoes and reverberations. Sound localization and robot navigation experiments were conducted. The results show that the robot is capable of localizing sounding objects in a reverberant environment and approaching the objects without collisions, even when the objects were behind obstacles. Environment flexibility and error robustness of the system were discussed as well.

Sound localization in reverberant environment based on the model of the precedence effect

This paper presents a model-based method for sound localization of concurrent and continuous speech sources in reverberant environment. A new algorithm adopted from the echo-avoidance model of the precedence effect was used to detect the echo-free onsets by specifying a generalized pattern of impulse response. Fine structure time differences were calculated from the zero-crossing points in different microphones. They were integrated into an azimuth histogram by the restrictions between them. Two sound sources were localized in both an anechoic chamber and a normal room which has walls, floor, and ceiling made of concrete. The time segment needed for localization was 0.5-2 s and the accuracy was a few degrees in both environments.

A biomimetic system for localization and separation of multiple sound sources

We present a system for sound source localization and separation inspired by the auditory mechanisms of biological systems. The system consists of three omni-directional microphones, banks of band-pass filters and a personal computer with a digital signal processor (DSP). Each microphone is set at a vertex of an equilateral triangle with a side length of 13.5 cm. First temporal disparities between microphones are detected by each band-pass filtered signal with some time duration (e.g. 30 sec.). It uses the onsets of a signal which are not corrupted with other sound sources including sound reflected by walls, etc. From the estimated azimuth of sound sources, we can calculate the time differences of sound arriving from different microphones. Each sound source is separated from each microphone signal using an inverse filter designed with this time difference. Experiments were carried out in an anechoic chamber and an empty room using two sound sources located with azimuth offset of 38 deg. We used a radio weather forecast by a male announcer and a radio talk show by a male and a female hosts as our sound sources. We localized each sound source with an error of less than 3 deg. and the separation of each sound source had a quality of 25 dB attenuation of each original sound. >

Efficient approximation of neural filters for removing quantum noise from images

In this paper, efficient filters are presented that approximate neural filters (NFs) that are trained to remove quantum noise from images. A novel analysis method is proposed for making clear the characteristics of the trained NF. In the proposed analysis method, an unknown nonlinear deterministic system with plural inputs such as the trained NF can be analyzed by using its outputs when the specific input signals are input to it. The experiments on the NFs trained to remove quantum noise from medical and natural images were performed. The results have demonstrated that the approximate filters, which are realized by using the results of the analysis, are sufficient for approximation of the trained NFs and efficient at computational cost.

A Simple Neural Network Pruning Algorithm with Application to Filter Synthesis

This paper describes an approach to synthesizing desired filters using a multilayer neural network (NN). In order to acquire the right function of the object filter, a simple method for reducing the structures of both the input and the hidden layers of the NN is proposed. In the proposed method, the units are removed from the NN on the basis of the influence of removing each unit on the error, and the NN is retrained to recover the damage of the removal. Each process is performed alternately, and then the structure is reduced. Experiments to synthesize a known filter were performed. By the analysis of the NN obtained by the proposed method, it has been shown that it acquires the right function of the object filter. By the experiment to synthesize the filter for solving real signal processing tasks, it has been shown that the NN obtained by the proposed method is superior to that obtained by the conventional method in terms of the filter performance and the computational cost.

Neural models of brightness perception and retinal rivalry in binocular vision

In binocular fusion, pairs of left and right stimuli yielding the same brightness perception constitute an equibrightness curve in a coordinate system whose ordinate and abscissa correspond to the left and right stimulus strengths. A neural network model is presented to elucidate the characteristics of the curve. According to the model, Fechners paradox is due to the threshold characteristics of the neuron. If the shapes or movements are radically different between the left and right stimuli, the retinal rivalry is caused. That is, only the left stimulus is perceived at one moment and the right stimulus at another moment. The period of left or right eye dominance alternates randomly from time to time. The distribution of the period is approximate to the gamma distribution. In order to account for this fact, a neural network model is proposed, which consists of a pair of neurons receiving inputs with stochastic fluctuations. The computer simulation was carried out with satisfactory results. The model of retinal rivalry is integrated with that of brightness perception.

Fast connected-component labeling based on sequential local operations in the course of forward raster scan followed by backward raster scan

Presents a fast algorithm for labeling connected components in binary images based on sequential local operations. A one-dimensional table, which memorizes label equivalences, is used for uniting equivalent labels successively during the operations in forward and backward raster directions. The proposed algorithm has a desirable characteristic: the execution time is directly proportional to the number of pixels in connected components in an image. By comparative evaluations, it has been shown that the efficiency of the proposed algorithm is superior to those of the conventional algorithms.

Building ears for robots: Sound localization and separation

This paper describes our research on bio-mimetic robot audition. Among the many binaural and monaural sound localization cues in the human auditory system, the interaural time difference cue is selected as it can easily be obtained by omnidirectional microphones. We have used a three-microphone system to remove the anterior-posterior ambiguity which occurs in two-microphone (or ear) systems. The echo-avoidance model of the precedence effect is used to cope with the echoes and reverberations of real environments. We mimicked the cocktail party effect by perceptual grouping of continuous components according to the spatial information obtained by the sound localization method. A wheel-based mobile robot equipped with an auditory system was developed. The auditory system has two sound processing parts. One is a DSP-based realtime system; the other is an off-line system composed of remote computers. Experiments of localizing and separating multiple sound sources and robot navigation were conducted to demonstrate the systems ability and potential applications.