Masahiro YASUDA Soh YOSHIDA Mitsuji MUNEYASU
Methods that embed data into printed images and retrieve data from printed images captured using the camera of a mobile device have been proposed. Evaluating these methods requires printing and capturing actual embedded images, which is burdensome. In this paper, we propose a method for reducing the workload for evaluating the performance of data embedding algorithms by simulating the degradation caused by printing and capturing images using generative adversarial networks. The proposed method can represent various captured conditions. Experimental results demonstrate that the proposed method achieves the same accuracy as detecting embedded data under actual conditions.
Hojun SHIMOYAMA Soh YOSHIDA Takao FUJITA Mitsuji MUNEYASU
Recent character detectors have been modeled using deep neural networks and have achieved high performance in various tasks, such as text detection in natural scenes and character detection in historical documents. However, existing methods cannot achieve high detection accuracy for wooden slips because of their multi-scale character sizes and aspect ratios, high character density, and close character-to-character distance. In this study, we propose a new U-Net-based character detection and localization framework that learns character regions and boundaries between characters. The proposed method enhances the learning performance of character regions by simultaneously learning the vertical and horizontal boundaries between characters. Furthermore, by adding simple and low-cost post-processing using the learned regions of character boundaries, it is possible to more accurately detect the location of a group of characters in a close neighborhood. In this study, we construct a wooden slip dataset. Experiments demonstrated that the proposed method outperformed existing character detection methods, including state-of-the-art character detection methods for historical documents.
Ryota HIGASHIMOTO Soh YOSHIDA Takashi HORIHATA Mitsuji MUNEYASU
Noisy labels in training data can significantly harm the performance of deep neural networks (DNNs). Recent research on learning with noisy labels uses a property of DNNs called the memorization effect to divide the training data into a set of data with reliable labels and a set of data with unreliable labels. Methods introducing semi-supervised learning strategies discard the unreliable labels and assign pseudo-labels generated from the confident predictions of the model. So far, this semi-supervised strategy has yielded the best results in this field. However, we observe that even when models are trained on balanced data, the distribution of the pseudo-labels can still exhibit an imbalance that is driven by data similarity. Additionally, a data bias is seen that originates from the division of the training data using the semi-supervised method. If we address both types of bias that arise from pseudo-labels, we can avoid the decrease in generalization performance caused by biased noisy pseudo-labels. We propose a learning method with noisy labels that introduces unbiased pseudo-labeling based on causal inference. The proposed method achieves significant accuracy gains in experiments at high noise rates on the standard benchmarks CIFAR-10 and CIFAR-100.
Mitsuji MUNEYASU Katsuya KONDO
Hiroshi Tsutsui Mitsuji Muneyasu
Soh YOSHIDA Takahiro OGAWA Miki HASEYAMA Mitsuji MUNEYASU
Video reranking is an effective way for improving the retrieval performance of text-based video search engines. This paper proposes a graph-based Web video search reranking method with local and global consistency analysis. Generally, the graph-based reranking approach constructs a graph whose nodes and edges respectively correspond to videos and their pairwise similarities. A lot of reranking methods are built based on a scheme which regularizes the smoothness of pairwise relevance scores between adjacent nodes with regard to a user's query. However, since the overall consistency is measured by aggregating only the local consistency over each pair, errors in score estimation increase when noisy samples are included within query-relevant videos' neighbors. To deal with the noisy samples, the proposed method leverages the global consistency of the graph structure, which is different from the conventional methods. Specifically, in order to detect this consistency, the propose method introduces a spectral clustering algorithm which can detect video groups, in which videos have strong semantic correlation, on the graph. Furthermore, a new regularization term, which smooths ranking scores within the same group, is introduced to the reranking framework. Since the score regularization is performed by both local and global aspects simultaneously, the accurate score estimation becomes feasible. Experimental results obtained by applying the proposed method to a real-world video collection show its effectiveness.
Mitsuji MUNEYASU Yumi WAKASUGI Ken'ichi KAGAWA Kensaku FUJII Takao HINAMOTO
A multiple channel active noise control (ANC) system with several secondary sources, error sensors, and reference sensors has been used for complicated noise fields. Centralized multiple channel ANC systems have been proposed, however implementation of such systems becomes difficult according to increase of control points. Distributed multiple channel ANC systems which have more than a controller are considered. This paper proposes a new implementation of distributed multiple channel ANC systems based on simultaneous equations methods. In the proposed algorithm, communications between controllers are permitted to distribute the computational burden and to improve the performance of noise reduction. This algorithm shows good performances for noise cancellation and tracking of changes in the error paths.
Takuto YOSHIOKA Kana YAMASAKI Takuya SAWADA Kensaku FUJII Mitsuji MUNEYASU Masakazu MORIMOTO
In this paper, we propose a step size control method capable of quickly canceling acoustic echo even when double talk continues from the echo path change. This method controls the step size by substituting the norm of the difference vector between the coefficient vectors of a main adaptive filter (Main-ADF) and a sub-adaptive filter (Sub-ADF) for the estimation error provided by the former. Actually, the number of taps of Sub-ADF is limited to a quarter of that of Main-ADF, and the larger step size than that applied to Main-ADF is given to Sub-ADF; accordingly the norm of the difference vector quickly approximates to the estimation error. The estimation speed can be improved by utilizing the norm of the difference vector for the step size control in Main-ADF. We show using speech signals that in single talk the proposed method can provide almost the same estimation speed as the method whose step size is fixed at the optimum one and verify that even in double talk the estimation error, quickly decreases.
Mitsuji MUNEYASU Shuhei ODANI Yoshihiro KITAURA Hitoshi NAMBA
On the use of a surveillance camera, there is a case where privacy protection should be considered. This paper proposes a new privacy protection method by automatically degrading the face region in surveillance images. The proposed method consists of ROI coding of JPEG2000 and a face detection method based on template matching. The experimental result shows that the face region can be detected and hidden correctly.
Kensaku FUJII Yoshihisa NAKATANI Mitsuji MUNEYASU
This paper proposes a new method to reduce sinusoidal noise components whose frequencies are known. The new method is based on the simultaneous equations technique. The technique does not require the secondary path filter: thereby the automatic recovering of the noise reduction effect deteriorated by secondary path changes becomes possible. This paper also presents computer simulation results to examine the performance of the new method.
Takamasa FUJII Soh YOSHIDA Mitsuji MUNEYASU
In video search reranking, in addition to the well-known semantic gap, the intent gap, which is the gap between the representation of the users' demand and the real search intention, is becoming a major problem restricting the improvement of reranking performance. To address this problem, we propose video search reranking based on a semantic representation by multiple tags. In the proposed method, we use relevance feedback, which the user can interact with by specifying some example videos from the initial search results. We apply the relevance feedback to reduce the gap between the real intent of the users and the video search results. In addition, we focus on the fact that multiple tags are used to represent video contents. By vectorizing multiple tags associated with videos on the basis of the Word2Vec algorithm and calculating the centroid of the tag vector as a collective representation, we can evaluate the semantic similarity between videos by using tag features. We conduct experiments on the YouTube-8M dataset, and the results show that our reranking approach is effective and efficient.
Mitsuji MUNEYASU Osamu HISAYASU Kensaku FUJII Takao HINAMOTO
A simultaneous equations method is one of active noise control algorithms without estimating an error path. This algorithm requires identification of a transfer function from a reference microphone to an error microphone containing the effect of a noise control filter. It is achieved by system identification of an auxiliary filter. However, the introduction of the auxiliary filter requires more number of samples to obtain the noise control filter and brings a requirement of some undesirable assumption in the multiple channel case. In this paper, a new simultaneous equations method without the identification of the auxiliary filter is proposed. By storing a small number of input signals and error signals, we avoid this identification. Therefore, we can reduce the number of samples to obtain the noise control filters and can avoid the undesirable assumption. From simulation examples, it is verified that the merits of the ordinary method is also retained in the proposed method.
Masayoshi NAKAMOTO Kohei SAYAMA Mitsuji MUNEYASU Tomotaka HARANO Shuichi OHNO
For copyright protection, a watermark signal is embedded in host images with a secret key, and a correlation is applied to judge the presence of watermark signal in the watermark detection. This paper treats a discrete wavelet transform (DWT)-based image watermarking method under specified false positive probability. We propose a new watermarking method to improve the detection performance by using not only positive correlation but also negative correlation. Also we present a statistical analysis for the detection performance with taking into account the false positive probability and prove the effectiveness of the proposed method. By using some experimental results, we verify the statistical analysis and show this method serves to improve the robustness against some attacks.
Mitsuji MUNEYASU Ken'ichi KAGAWA Kensaku FUJII Takao HINAMOTO
For multiple-channel active noise control (ANC) systems, distributed systems consisting of more than one controller are useful. In this paper, we propose a performance improvement algorithm for the distributed multiple-channel ANC system based on the simultaneous equations method. In the proposed algorithm, no estimation of error paths is required. This algorithm can provide good performance in canceling primary noises with auto-/cross-correlations and achieve stable noise reduction under a change of the error paths.
Kensaku FUJII Kenji KASHIHARA Mitsuji MUNEYASU Masakazu MORIMOTO
In this paper, we propose a method capable of shortening the distance from a noise detection microphone to a loudspeaker, which is one of important issues in the field of active noise control (ANC). In the ANC system, the secondary noise provided by the loudspeaker is required arriving at an error microphone simultaneously with the primary noise to be cancelled. However, the reverberation involved in the secondary path from the loudspeaker to the error microphone increases the secondary noise components arriving later than the primary noise. The late components are not only invalid for canceling the primary noise but also impede the cancellation. To reduce the late components, the distance between the noise detection microphone and the loud speaker is generally extended. The proposed method differently reduces the late components by forming the noise control filter, which produces the secondary noise, with the cascade connection of a non-recursive and a recursive filters. The distance can be thus shortened. On the other hand, the recursive filter is required to work stably. The proposed method guarantees the stable work by forming the recursive filter with the lattice filter whose coefficients are restricted to less than unity.
Kensaku FUJII Shigeyuki HASHIMOTO Mitsuji MUNEYASU
This paper presents a frequency domain simultaneous equations method capable of automatically recovering noise reduction effect degraded by secondary path changes. The simultaneous equations method has been studied, first in time domain. Accordingly to the study, in the time domain, the simultaneous equations method requires an additional filter and a system identification circuit used for transforming the solution of the simultaneous equations into the coefficients of noise control filter, which increase the processing cost. To reduce the processing cost, this paper studies on the application of a frequency domain processing technique, the cross spectrum method, to the simultaneous equations method. By directly applying the equation defining the cross spectrum method to the solution, the additional filter becomes unnecessary. In addition, the system identification circuit is replaced with the inverse Fourier transform. Thereby, the processing cost drastically decreases. This paper also presents simulation results to confirm that the proposed method can automatically recover the noise reduction effect degraded by a path change and provides much higher convergence speed than that of the filtered-x NLMS algorithm with the perfectly modeled secondary path filter.
Mitsuji MUNEYASU Kouichiro ASOU Yuji WADA Akira TAGUCHI Takao HINAMOTO
This paper presents a new implementation of fuzzy filters for edge-preserving smoothing of an image corrupted by impulsive and white Gaussian noise. This filter structure is expressed as an adaptive weighted mean filter that uses fuzzy control. The parameters of this filter can be adjusted by learning. Finally, simulation results demonstrate the effectiveness of the proposed technique.
Kensaku FUJII Mitsuji MUNEYASU Takao HINAMOTO Yoshinori TANAKA
The sub-recursive least squares (sub-RLS) algorithm estimates the coefficients of adaptive filter under the least squares (LS) criterion, however, does not require the calculation of inverse matrix. The sub-RLS algorithm, based on the different principle from the RLS algorithm, still provides a convergence property similar to that of the RLS algorithm. This paper first rewrites the convergence condition of the sub-RLS algorithm, and then proves that the convergence property of the sub-RLS algorithm successively approximates that of the RLS algorithm on the convergence condition.
Liang LI Akira ASANO Chie MURAKI ASANO Mitsuji MUNEYASU Yoshiko HANADA
A method of estimating dual primitives in a textural image is proposed. This method is based on the Primitive, Grain, and Point Configuration (PGPC) texture model, which regards a texture as an arrangement of grains derived from one or a few primitives. Appropriate primitives can be represented by morphological structuring elements estimated from a texture. Conventional primitive estimation methods estimate only one primitive from each textural image. However, they do not work well on textural images that contain more than one basic structure, since two or more types of grain cannot be generated from only one primitive. The proposed method simultaneously estimates two optimal structuring elements of a texture. The experimental results show that the proposed method provides more representative estimations than the conventional method.
Hiroyuki OKUNO Yoshiko HANADA Mitsuji MUNEYASU Akira ASANO
In this paper we propose an unsupervised method of optimizing structuring elements (SEs) used for impulse noise reduction in texture images through the opening operation which is one of the morphological operations. In this method, a genetic algorithm (GA), which can effectively search wide search spaces, is applied and the size of the shape of the SE is included in the design variables. Through experiments, it is shown that our new approach generally outperforms the conventional method.