Traffic sign recognition system is an intelligent vision system to recognize traffic signs on the road. This paper describes a new approach on red color thresholding for traffic sign recognition system. To extract red color from an image, a thresholding method based on the CIE-RGB chromaticity diagram was proposed. A g-r histogram of the image was developed by subtracting the normalized red of RGB color from the normalized green one. A threshold value is selected automatically by analyzing the g-r histogram. To evaluate the proposed method, comparison to the other red color thresholding methods was made using tested images taken from real environments with varying illuminations and backgrounds. From the experiment, our proposed method resulted in the thresholding quality with average score of 0.630 and average execution time of 0.038 seconds which is the best performance among the others.

View full abstract

In this paper a new method to detect road signs from color road images and to discriminate them is proposed. This problem has been recognized as an important research field of ITS(Intelligent Transportation Systems). Proposed Self-Organizing Map is constructed by using three kinds of indices, feature of color rate, shape feature coefficient and texture feature coefficient. Experimental results show this method is sufficiently efficient for practical use.

View full abstract

In recent years, many different image features have been used for face recognition. The Gabor wavelet feature is the most widely used image feature in face recognition systems because its recognition rate is superior to that of Eigenface systems. However, it remains unclear as to whether Gabor wavelet features are indeed the best wavelet features for face recognition. In this paper, we extract image features of facial images from various wavelet transforms (e.g., Haar, French hat, Mexican hat, Daubechies, Coiflet, Symlet, and O-spline) and evaluate their face recognition performance. We also compare the recognition performance of fixed-and adaptive-scale wavelet features. We rotate wavelets, extract features and evaluate their recognition performance. The results demonstrate that the performance of Haar wavelet is superior to that of the Gabor wavelet.

View full abstract

In order to realize a position-based service such as transferring useful information to users located in an area, several infrared-used systems have been proposed. Without making significant changes to the configuration, we added an upload function that helps to send user information to the system. A user can send his/her data by using a handheld device that has a reflective sheet and a liquid crystal on its front. The infrared in the system illuminates in the area where the user's device is lying. When the liquid crystal is clear, the reflective infrared magnitude is large. In contrast, the infrared magnitude is comparatively small when the crystal is dark. An infrared camera placed near the infrared projector recognizes the difference in the magnitude. This difference acts as a communication signal from the user to the system. It is necessary to determines the static threshold for distinguishing these two conditions; however, this is difficult because the difference also depends on the position and angle of the user's device. Therefore, the system we have proposed in this research first recognizes the sync bit sent from the user's device by using a Bayesian network. By means of this recognition process, the threshold can then be determined dynamically. Moreover, our experimental results show that the proposed method can easily overcome previously proposed machine learning methods.

View full abstract

A restoration algorithm based on the Image Inpainting technique is proposed for image restoration of Halo regions around light sources in digital night images. The proposed algorithm consists of two processes; first, each diffused halo region around light sources is recognized as a fuzzy set based on “degree of illuminant”. Secondly, the background pixels which are pale because of the insufficient light are updated with the information from the pixels surrounding the halo region. Moreover, a tone control filter for the halo region is proposed to improve the performance of the image restoration process. Experimental results on night images taken by a commercially available digital camera show that the proposed algorithm produces a halo region contracted by an average of 30% in size when compared to the input image using a 3 dimension “degree of illuminant” graph. The results are produces in competitive computational time.

View full abstract

In the field of image processing, it is very important in the object recognition to perform the color extraction processing by deciding the fittest threshold. In general, the threshold of color information dynamically changes according to the locations and lighting conditions in the real environment including various noises. Therefore, it is not easy for human to obtain the threshold values of a target object in the real environment. In this paper, we propose a threshold tuning method that is able to extract only the target area without noises by the threshold values of an ellipse target object selected by human from color static image obtained by an omnidirectional camera. In this method, the tuning of color extraction threshold is performed by genetic algorithm (GA) to search the suitable color threshold value according to the color information in the area selected by human. We also report the results of experiments performed by comparing the performance with the threshold value before and after GA search. By these experimental results, we confirmed that the proposed color extraction method by GA has better and faster performance than that of human.

View full abstract

Recently, Content-based Image Retrieval (CBIR) which evaluates the relevance of images derived by image features is being widely investigated, aiming at retrieval from large database of unlabeled images. Most commonly in CBIR, “similarity”, which generally does not have common or objective definition is used. Naturally, “similarity” can be defined in various ways and it differs according to the user's subjectivity. Because of this ambiguity of similarity, realization of retrieval suited to the user's similarity criterion is needed. This work proposes a framework for improving the similarity evaluation of images according to the user's demand, by optimizing the parameters in the relevance evaluation algorithm according to the criteria defined by the user. In this framework, we define a criteria which is a function of the retrieved result, and its parameters are optimized by using Particle Swarm Optimization (PSO). As a system to evaluate the effectiveness of the proposed framework, we used a CBIR system of binary images which is based on the matching of image contour features. The evaluation function J based on the user's rating of the retrieval result was defined and the parameters in the relevance evaluation algorithm were optimized. In the first experiment, the parameters were optimized in the proposed framework for a given evaluation function J using PSO. There, it was found that retrieval sets more suited according to the evaluation function was obtained after the optimization. In the second experiment, two different criteria were defined. When the system was optimized for each different criterion, it was found that the final parameters were different for different requirements, and that retrieved results were also different, each adapting to the specific requirement. As a whole, it was found that CBIR systems suited to the user's subjective similarities can be obtained by using the proposed framework, defining criteria J and optimizing the parameters accordingly.

View full abstract

Content-based image retrieval has become an increasingly popular field of research in computer vision, however the problem of the semantic gap between low-level features, such as color, shape and texture, and high-level semantics remains unsolved. Recent research in content-based image retrieval has shifted to an interactive process that considers the user as a part of the retrieval process. In order to realize this iterative process, relevance feedback is introduced into content-based image retrieval. The purpose of the present study is, therefore, to develop a query-by-sketch image retrieval method for reducing the semantic gap between low-level features and high-level semantics by adopting relevance feedback. In the proposed method, users' sketches play an important role in reducing the semantic gap by relevance feedback. This method is applied to 6,500 images in Corel Photo Gallery. Experimental results show that the image retrieval system with the relevance feedback is superior to the image retrieval system without the relevance feedback.

View full abstract

Intelligent surveillance systems capable of discriminating pedestrians in the street are one of the main application areas of computer vision. This paper proposes a method to discriminate pedestrian images into several classes by using pedestrian shape features and artificial neural networks. To overcome the difficulty of pedestrian identification due to shape variation over time, several video-image processing and intelligent discrimination methods were adopted and developed. At the front end, image and video processing was performed to separate the background from the foreground images. The pedestrian shape features were extracted by Fourier transform, and then feed-forward neural networks with back-propagation learning algorithms were employed to discriminate among several classes of the moving pedestrian images, i.e., pedestrian, cyclist, or other non-pedestrian objects. The experimental results demonstrated the capability of the proposed system to discriminate pedestrians in a real life pedestrian environment. On average, discrimination accuracy was achieved in 82% and 87% using the complex number and the centroid distance function method, respectively.

View full abstract

The goal of this research is to develop a real world search system for everyday objects using position sensors on those objects. The aim of this paper is to realize a search method which can use past landmark near by the searching object as an indicator for the search. By using a camera, a user suggests a place where the searching object were in his/her memory, using a landmark. There are two problems for realizing such a landmark indication search. The first problem is the losing landmark problem. It is the problem which the user can not find the landmark at the location he/she remembers. The second one is the confusing landmarks problem which the user might select the wrong landmark when landmarks have similar features. To solve these two problems, we introduce Atumeye frames. Atumeye frames are based on the idea of landmark memories of the user that when the user memorizes a landmark, he/she also memorizes relative physical relationship between the landmark and other landmarks around. Brownie stores the location of an object, and the physical relationship around the landmark of the object as one Atumeye frame. The physical relationship is decided by existences of overlap regions of landmark objects in the video camera, and the overlap region is calculated from the 3D position data. To solve the losing landmark problem, Brownie estimates Atumeye frames when the problem occurs. To solve the confusing landmarks problem, when the user runs the search with the landmark, Brownie runs another search using other landmarks which have similar relative physical relationship. The result of evaluation indicates the average detection rate of real world search is 76.7% with Atumeye frames.

View full abstract

A nonlinear digital filter named as an extended component-separating filter (an ECS filter for short) is proposed for removing additive random noise from signals with abrupt changes, such as images and speech, while restoring the signal waveform as precise as possible. This filter corresponds to an extended version of nonlinear filters, such as an ε-filter and a component-separating filter (a CS filter for short) and realizes higher performance than these nonlinear filters. Especially, the ECS filter is effective for both a step-like signal and a continuously changing signal, while the ε-filter and the CS filter are effective for either of them. Since the filter structure corresponds to a layered neural network, the ECS filter can be optimized by a training method as the back-propagation algorithm using a training signal. In computer simulations, this filter is shown to be quantitatively effective for both step-like signals and continuous ones, giving totally small mean square error for both signals. Moreover, this filter is shown to be quite effective for face image processing to make the skin look smooth and beautified.

View full abstract

A 3D model is used for an expression of product model and object recognition besides the simulation. The applied field is wide, and the importance is very high. However, an extensive cost is still needed for the creation of 3D model. Therefore, there are various researches on generating the 3D model quickly, easily, and at a low cost. The general technique has not been established. Then, in this research, we try to establish the technique of automatic creation of the 3D model by the photogrammetry from a digital animation filmed by digital video camera to satisfy the above-mentioned needs. Especially, we aim to solve the problem of the previous research. That is to improve the accuracy and speed for tracking feature points on an object for 3D modeling by the optical flow. Concretely, there are two subjects to decrease the processing speed and to improve the accuracy of movement and rotation in template matching method. We solved them using the optical flow by the Harris operator.

View full abstract