Cross-modal retrieval with noisy labels has attracted much attention. This state-of-the-art method trains a network to increase weights for clean labels in the loss. However, we have found that the network is eventually overfitted to the remaining noisy labels as training progresses. Motivated by this finding, this paper proposes a method called Label Correction using Network prediction based on Memorization Effects (LCNME) to correct noisy labels. This is unlike the state-of-the-art method, which leaves noisy labels on training. We assume that noisy labels are irrelevant to data features and realize label correction using predicted labels (obtained by network prediction) instead of given labels. However, because of memorization effects (the property whereby the network first learns clean labeled data then learns noisy labeled data), predicted labels are contaminated by noisy labels from the certain epoch called the change epoch. Although the change epoch is unknown in advance, we find that it can be identified by observing the loss of the noisy validation set. Using the change epoch, predicted labels can be generated without being affected by noisy labels. Extensive experiments show that LCNME accurately corrects noisy labels and achieves better cross-modal retrieval than existing methods.

For a jar test in water purification, it is necessary to develop a video prediction method to assist in coagulant dosage decisions. However, such a method has not been established; video prediction methods in other fields require a large amount of training videos. This paper proposes a video prediction method based on multi-resolution echo state networks (ESN) for a jar test. Because the proposed method is constructed based on ESN, we succeed in long-term video prediction even if the amount of training videos is small. Specifically, we develop a new model that performs ESN while observing the input video at different resolutions and effectively fuses the prediction results. The multi-resolution approach enables us to acquire both rough shape changes and fine pattern features, resulting in accurate prediction. In fusion of the multiple prediction results, our method can calculate optimal weights by the least squares method. Experimental results for a jar test confirm that our method outperformed existing video prediction methods based on deep neural networks. In addition, the results on the sky timelapse dataset show the versatility of our method.

Feature representation of oleaginous yeast images is essential for non-invasively observing their cultivation status. However, conventional methods are difficult to capture the visual characteristics of oleaginous yeasts at the extremely local spatial regions. To tackle this problem, this paper proposes micro-spatial attention with sparse constraints for self-supervised learning. Specifically, we newly design a micro-spatial attention module to focus on the extremely local spatial regions for feature representation. Unlike the existing spatial attention module, which requires the guide information (e.g., labels and textual queries) to focus on specific regions, our proposed module employs a sparse constraint. By combining the sparse constraint with self-supervised learning, we can automatically focus on essential regions, which appear in extremely local regions. This improves the feature representation power for oleaginous yeast images with micro-spatial contexts. Experiments on real Lipomyces starkeyi datasets, a specific type of oleaginous yeasts, confirm the superiority of our method. Moreover, our method realized clustering of oleaginous yeast images, enabling us to non-invasively observe the cultivation trends over time.

Learning of the convolutional neural network (CNN) from noisy labeled samples is a crucial problem, and many studies have been conducted. Although the state-of-the-art method called Joint training with Co-Regularization (JoCoR) has achieved high performance, it is a still challenging problem to accurately classify samples with asymmetric noise, i.e., wrong labels between similar classes (e.g., CAT ↔ DOG). In this paper, we newly found that there is a difference between the prediction norm for clean labeled samples and that for noisy labeled samples. In addition, we found that there is a positive correlation between the prediction norm difference and the classification accuracy for the CIFAR-10 dataset. Therefore, we hypothesize that the discriminative power would be improved if we increase the prediction norm difference. Based on this hypothesis, a novel method for learning CNN from noisy labeled samples is proposed. Specifically, we take JoCoR as the base architecture and weight the loss function of JoCoR to make the prediction norm difference large. Experimental results for the CIFAR-10 dataset suggest that our hypothesis is correct. The classification accuracy by the proposed method is higher than some state-of-the-art methods including JoCoR for samples with asymmetric noise as well as those with symmetric noise (i.e., wrong labels between other classes).

Abstract Electroencephalography (EEG) is a method for recording electrical activities arising from the cortical surface of the brain, which has found wide applications not just in clinical medicine, but also in neuroscience research and studies of Brain‐Computer Interface (BCI). However, EEG recordings often suffer from distortions due to artefactual components that degrade the true EEG signals. Artefactual components are any unwanted signals recorded in the EEG spectrum that originate from sources other than the neurophysiological activity of the human brain. Examples of the origin of artefactual components include eye blinking, facial or scalp muscles activities, and electrode slippage. Techniques for automated artefact removal such as Wavelet Transform and Independent Component Analysis (ICA) have been used to remove or reduce the effect of artefactual components on the EEG signals. However, detecting or identifying the signal artefacts to be removed presents a great challenge, as EEG signal properties vary between individuals and age groups. Techniques that rely on some arbitrarily defined threshold often fail to identify accurately the signal artefacts in a given dataset. In this study, a method is proposed using unsupervised machine learning coupled with Wavelet‐ICA to remove EEG artefacts. Using Density‐Based Spatial Clustering of Application with Noise (DBSCAN), a classification accuracy of 97.9% is achieved in identifying artefactual components. DBSCAN achieved excellent and robust performance in identifying artefactual components during the Wavelet‐ICA process, especially in consideration of the low‐density nature of typical artefactual signals. This new hybrid method provides a scalable and unsupervised solution for automated artefact removal that should be applicable for a wide range of EEG data types.

. By identifying the tweet communities that have high importance measures, it is possible for users to easily find important information about the coronavirus disease (COVID-19). Specifically, we first construct a community network, whose nodes are tweet communities obtained by applying a community detection method to a tweet network. The community network is constructed based on textual similarities between tweet communities and sizes of tweet communities. Second, we apply algorithms for calculating centrality to the community network. Because the obtained centrality is based on tweet community sizes as well, we call it the importance measure in distinction to conventional centrality. The importance measure can simultaneously evaluate the importance of topics in the entire data set and occupancy (or dominance) of tweet communities in the network structure. We conducted experiments by collecting Japanese tweets about COVID-19 from March 1, 2020 to May 15, 2020. The results show that the proposed method is able to extract keywords that have a high correlation with the number of people infected with COVID-19 in Japan. Because users can browse the keywords from a small number of central tweet communities, quick and easy understanding of important information becomes feasible.

Many measurement methods have been proposed for use in automated production. Existing methods for measuring three-dimensional surface height data operate by projecting fringe patterns onto a target object and capturing images of them. However, these methods do not work well for cutlery, such as table forks and spoons, because specular reflection causes exposure change on the cutlery surface. To overcome this difficulty, this paper presents a measurement method for cutlery with specular reflection. The proposed method is based on a well-known measurement technique, the gray coding method, which is robust to exposure change because the captured fringe patterns are handled by discretizing their values. Furthermore, the gray coding method is improved by introducing hue-channel information into the captured fringe patterns. The hue-channel information is more robust to exposure change than brightness, which is used in existing methods. Experimental results for real cutlery show that the proposed hue-based gray coding method enables more robust measurement than existing methods.

Brain electrical activity recordings by electroencephalography (EEG) are often contaminated with signal artifacts. Procedures for automated removal of EEG artifacts are frequently sought for clinical diagnostics and brain-computer interface applications. In recent years, a combination of independent component analysis (ICA) and discrete wavelet transform has been introduced as standard technique for EEG artifact removal. However, in performing the wavelet-ICA procedure, visual inspection or arbitrary thresholding may be required for identifying artifactual components in the EEG signal. We now propose a novel approach for identifying artifactual components separated by wavelet-ICA using a pretrained support vector machine (SVM). Our method presents a robust and extendable system that enables fully automated identification and removal of artifacts from EEG signals, without applying any arbitrary thresholding. Using test data contaminated by eye blink artifacts, we show that our method performed better in identifying artifactual components than did existing thresholding methods. Furthermore, wavelet-ICA in conjunction with SVM successfully removed target artifacts, while largely retaining the EEG source signals of interest. We propose a set of features including kurtosis, variance, Shannon's entropy, and range of amplitude as training and test data of SVM to identify eye blink artifacts in EEG signals. This combinatorial method is also extendable to accommodate multiple types of artifacts present in multichannel EEG. We envision future research to explore other descriptive features corresponding to other types of artifactual components.

The Wavelet Transform (WT) in JPEG 2000 is using a 'separable' lifting structure, where the one-dimensional (1D) transform is put into multidimensional image signal of its spatial and temporal dimensions. A 'non-separable' three-dimensional (3D) structure as the existing method is used to minimize its lifting steps. The 'non-separable' 3D structure in the (5,3) type transform for lossless coding is proved to reduce the rounding noise inside it. However, in the (9,7) type transform for lossy coding, the rounding noise inside the 'non-separable' 3D structure has increased. This paper proposed a new 'non-separable' two-dimensional (2D) structure for integer implementation of a four-dimensional (4D) quadruple lifting WT. Since the order of the original lifting step is preserved, the total amount of the rounding noise observed in pixel values of the decoded image is significantly reduced, and the lossy coding performance for 4D input signal is increased.

Phase information is ignored for almost all voice activity detection (VAD). To exploit full information in the original signal, this paper proposes a deep neural network (DNN) using magnitude and phase information (that is, phase aware DNN) to achieve better VAD performance. Mel-frequency cepstral coefficient (MFCC), power-normalized cepstral coefficients (PNCC), instantaneous frequency derivative (IF), baseband phase difference (BPD) and modified group delay cepstral coefficient (MGDCC) are used as magnitude and phase information. The proposed methods were evaluated using CENSREC-1-C database under noise condition. The results show that the phase aware DNN significantly outperforms the DNN using only magnitude information. For DNN-based classifier, the equal error rate (EER) was reduced from 23.70% of MFCC, to 20.43% of joint dual magnitude and single phase features (augmenting PNCC, MGDCC and IF), to 19.92% of joint dual phase and single magnitude feature features (augmenting PNCC, MGDCC and BPD). By combining joint dual magnitude and single phase features with joint dual phase and single magnitude features, the EER was reduced to 19.44%.

The super resolution technique plays an important role in displaying a conventional low resolution (LR) image on a recently developed high resolution (HR) display. In its process, some of enhancement techniques such as de-noising, de-blurring and tone-mapping are combined to increase visual quality. This paper introduces a subset noise bias compensation (NBC) to an LR-to-HR up-conversion process which is composed of tone-mapping and up-scaling. In our system, 1) an LR image is encoded and decoded with JPEG, 2) pixels in the decoded noisy image are tone-mapped, 3) observed pixel values are classifies into subsets and the bias of each subset is compensated with an optimally designed compensation value, and 4) the images are up-scaled with a state-of-the-art super-resolution technique. Improvement by the proposed method is confirmed with PSNR and SSIM evaluation measures, respectively.

Implementation issues of the separable 2D lifting wavelet transform has been discussed under the international standard JPEG 2000 scheme. Recently, the minimum lifting step structure of the transform was proposed introducing the multi-dimensional memory accessing. It can reduce latency in a parallel processing platform. However, its output signal is distorted due to truncation of multiplier coefficient values. Especially, a few coefficients are extremely sensitive to the truncation in this structure. Therefore a special treatment is required to these coefficients. In this paper, a tolerable (the maximum) truncation error is optimally assigned to each coefficient under a given cost function introducing the noise gain. It is applied to expressing coefficient values in the sum-of-power-of-two (SPT) format for reducing computational cost of multipliers of the transform.

This paper proposes a de-noising method especially for a biased noise. Even though a noise is assumed to have zero mean, an image processing such as tone mapping alters this property. Firstly, a precise estimation of the noise bias caused by a tone mapping function is derived considering discontinuity of a general form of the tone mapping. Secondly, a method of compensating the noise bias is introduced investigating estimation from observed noisy pixel values. Thirdly, the derived estimation is applied not only to compensate the noise bias, but also to reduce variance of the biased noise. Finally, significance of the proposed method is evaluated using HDR images in combination with a non-local mean filter.

We propose an integer tone mapping operator (TMO) for high dynamic range (HDR) images expressed in a floating-point data format. Two purposes are achieved by the proposed TMO. The first purpose is to implement a TMO with less memory space. The second purpose is to give an important step to realize a fixed-point TMO. The proposed TMO is available for HDR images in the OpenEXR format. The OpenEXR format has two numerical representations (the normalized number and the denormalized number) which are not in other HDR formats such as RGBE. These two numerical representations cause a problem in applying an integer TMO. The proposed method enables us to avoid the problem by using the intermediate format. Moreover, the exponent part and the mantissa part are processed separately as two integer numbers. As a result, an integer TMO with less numerical range is achieved by our method. The experimental results show that the proposed method can generate high-quality low dynamic range (LDR) images with less memory space.

This report proposes a two layer lossless coding system for high dynamic range (HDR) images expressed in a floating point data format. Its encoder outputs compressed image data in two layers. From the base layer, a standard low dynamic range (LDR) image is decoded. Merging with the enhancement layer, the original HDR image is decoded without any loss. We introduce a reversible logarithmic mapping to reduce bit depth of pixel values. We also introduce bit plane classification to divide data into two layers. It was confirmed that our method has better coding performance than existing methods especially at high bit rate coding of LDR images.

This report proposes a new tone mapping operation which is implemented in integer input and integer output. A tone mapping operation (TMO) generates a low dynamic range (LDR) image from a high dynamic range (HDR) image. Since pixel values of an HDR image are generally expressed in a floating point data format, e.g. RGBE, OpenEXR, a TMO is also implemented in floating point calculations in conventional approaches. However, it requires huge memory resources, even though a resulting LDR image is expressed in simple integer. We perform a TMO with integer input and integer output to reduce memory resources. It is experimentally confirmed with PSNR and contrast evaluations that the proposed method offers LDR images of visually high quality comparable to the conventional method.

This report reduces the total number of lifting steps in a three-dimensional (3D) double lifting discrete wavelet transform (DWT), which has been widely applied for analyzing volumetric medical images. The lifting steps are necessary components in a DWT. Since calculation in a lifting step must wait for a result of former step, cascading many lifting steps brings about increase of delay from input to output. We decrease the total number of lifting steps introducing 3D memory accessing for the implementation of low delay 3D DWT. We also maintain compatibility with the conventional 5/3 DWT defined by JPEG 2000 international standard for utilization of its software and hardware resources. Finally, the total number of lifting steps and rounding operations were reduced to 67 % and 33 %, respectively. It was observed that total amount of errors due to rounding operations in the lifting steps was also reduced.

The accurate detection of event-related potentials (ERPs) is of great importance to construct brain-machine interfaces (BMI) and constitutes a classification problem in which the appropriate selection of features from dense-array EEG signals and tuning of the classifier parameters are critical. In the present work, we propose a method for classifying single-trial ERPs using a combination of the Lifting Wavelet Transform (LWT), Support Vector Machines (SVM) and Particle Swarm Optimization (PSO). In particular, the LWT filters, the set of EEG channels and SVM parameters that maximize the classification accuracy are searched using PSO. We evaluate the method's performance through offline analyses on the datasets from the BCI Competitions II and III. The proposed method achieved in most cases a similar or higher classification accuracy than that achieved by other methods, and adapted wavelet basis functions and channel sets that match the time-frequency and spatial properties of the P300 ERP.

In this paper, a lossy data compression for a sparse histogram image signal is proposed. It is extended from an existing lossless coding which is based on a lossless histogram packing and a lossless coding. We introduce a lossy mapping, which has less computational load than the rate-distortion optimized Lloyd-Max quantization, and combine it with a lossless coding. It was confirmed that the proposed method attains higher performance in the rate-distortion plane than existing methods. This is because it can utilize histogram sparseness of images, and also its inverse mapping does not magnify quantization noise.

In this report, we propose a lossy compression algorithm which can utilize the `histogram sparseness' of pixel values of an input image signal. The bit depth of pixel values has becoming longer than conventional ones in high dynamic range images. Their histogram of pixel values tends to be `sparse', - not all the histogram bins are used. We extend an existing lossless coding based on the histogram packing to a lossy coding, introducing a non-uniform quantization especially for histogram sparse images. It was confirmed that the proposed method attains finer control of the bit rate and also higher SNR in the rate-distortion curve comparing to the existing method.

This report proposes an integer color transform for lossless coding of four color components of images. An existing color transform has a fixed set of coefficients and therefore it can't be adaptive to each image. We utilize eigen-vector of the covariance of the four components to increase data compaction performance. We also utilize fixed relation between two green components to simplify computational cost of the transform. It was confirmed that the proposed method reduces entropy rate by 0.4 bit comparing to the existing method.

In this report, we reduce the number of lifting steps of the two dimensional (2D) discrete wavelet transform (DWT) in JPEG 2000 by introducing a factorization of the separable 2D transfer function into some non separable 2D (NS-2D) functions. The order of the 1D lifting steps in vertical and in horizontal are properly replaced so that all of them are synthesized into NS-2D functions. As a result, the number of the lifting steps of the 2D 9/7 DWT is reduced to 75 [%] maintaining compatibility to the JPEG 2000. Since the number of the lifting steps is essentially proportional to a delay of output signals, it contributes to construct a low latency DWT under the assumption that direct accessing to a 2D intermediate memory is permitted.

This report theoretically derives a condition on the word length of coefficients and signals such that the irreversible 9-7 discrete wavelet transform (DWT) in JPEG 2000 becomes lossless for a constant valued (DC) input signal. Under the condition, the output signal contains no error in spite of rounding of coefficients and signals inside the DWT circuit. We have previously analyzed the condition for any value of DC signals. However, it was too strict and the word length under the condition was redundant for a practical case. In this report, we derive a new condition for a specific DC value and show that the minimum word length is halved under the new condition.

This paper proposes a “reversible” three-point Karhunen Loeve transform (KLT) for de-correlation of RGB color components of an image signal. It is composed of “reversible” two point rotation transforms which have singular points (SP) and their rotation angles change depending on correlation of an input signal. When the angle is close to SP, rounding errors inside the transform are magnified to huge amount. To avoid this problem, we introduce permutation “order” and “sign” of signals. We also make it clear that the proposed method can shift SP by 0, 90, 180 or 270 degrees so that distance between SP and the rotation angle is maximized. It was observed that the proposed method reduces variance of the error by 10 [%].

This report proposes a new transcoding between a lossless encoder and a lossy decoder for color image signals. It requires a reversible color transform (Rev.CT) with compatibility to an irreversible color transform (Irrv.CT). An existing Rev.CT guarantees lossless reconstruction of signals. However its compatibility to an Irrv.CT is not high enough for practical use. This is because a scaling is excluded from the Rev.CT for lossless coding. In this report, we embed the scaling into quantization in lossy decoding to realize both lossless coding and transcoding with high compatibility. It is experimentally confirmed that image quality is improved by more than 37.6 [dB]. We also evaluate its performance based on various criteria including, bit-extension, word-length of signals and coefficients.

In this report, permutation of order and sign of signals are introduced to avoid singular point problem of a reversible transform. When a transform is implemented in the lifting structure, it can be "reversible" in spite of rounding operations inside the transform. Therefore it has been applied to lossless coding of digital signals. However some coefficient values of the transform have singular points (SP). Around the SP, rounding errors are magnified to huge amount and the coding efficiency is decreased. In this report, we analyze the SP of a three point KLT for RGB color components of an image signal, and introduce permutation of order and sign of signals to avoid the SP problem. It was experimentally confirmed that the proposed method improved PSNR by approximately 15 [dB] comparing to the worst case.

In this report, we propose a reversible two dimensional (2D) discrete wavelet transform (DWT) compatible to the irreversible 9-7 DWT for lossy coding in the JPEG 2000. All the filters and scalings are factorized into lifting steps, and signals are rounded into integers, so that the proposed DWT becomes reversible and applicable to lossless coding. Furthermore, we factorize the separable 2D transfer function of the 2D DWT into non separable 2D functions to reduce the lifting steps and rounding operations. As a result, errors due to the rounding operations are reduced and the compatibility to the irreversible 9-7 DWT is improved.

In this paper, a new structure of three dimensional integer rotation transform is proposed. The number of rounding operations is reduced from nine to four by introducing multi input lifting branches. Furthermore, the best combination of order of input signals is determined to avoid singular points of multiplier coefficients in the lifting branches. As a result, variance of errors due to rounding of signals into integers is reduced by 42.9[%] at maximum and its compatibility to the conventional non integer rotation transform is improved.

This paper proposes a functionally layered video coding for robot vision network. Video signals from mobile robots are regularly transmitted to a remote robot. It uses the video to estimate locations of each of the mobile robots. As the need arises, a person browses the video to check the scenery. In such occasion, the functionally layered video coding regularly transmits the minimum components necessary for the motion estimation by the remote robot. For scenery browsing, only additional components are transmitted on demand. As a result, data volume to be transmitted is reduced in total. The proposed method decomposes the video into frequency bands and bit planes by means of the JPEG 2000. This report investigates the minimum components for the motion estimation in the robot to robot communication. It is confirmed that the proposed method can reduce total data volume of the robot vision network.

This paper proposes a new class of lifting wavelet transform which can guarantee losslessness of specific signals, e.g. white balance. The 5/3 wavelet transform composed of two lifting steps can reconstruct an input signal without any loss and has been utilized for lossless coding. The 9/7 wavelet contains two more lifting steps and two scaling pairs for effective lossy coding. However the losslessness is not guaranteed due to rounding of signal values and scaling coefficient values. This paper analyzes condition on word length (WL) and bit depth (BD) for the losslessness and proposes a new class of wavelet transform with "DC lossless " property which is a kind of specific losslessness. This can be utilized as a standard condition for algorithms or LSI processors to guarantee no error from the wavelet transform for white balance signals.

The 5/3 wavelet transform with double lifting steps in JPEG 2000 can reconstruct a signal without any loss. It has been utilized for lossless coding. The 9/7 wavelet transform contains two more lifting steps and scaling operations to improve performance for lossy coding. The loss is due to (1) quantization of band signals, (2) rounding of signals after scaling and (3) finite word length expression of scaling coefficients. This paper analyzes conditions on word length of coefficients and bit depth of rounded signals for no loss. It also proposes a new structure of lifting wavelet by changing order of the lifting step and the scaling. As a result, the rounding error is not scattered by the lifting steps and the error is minimized in mini-max sense.

This paper proposes how to detect water level from a video signal for use of automatic river surveillance. The water level is recognized as a boundary line between the land region and the water region. A conventional method uses a "vertical" edge detector to extract a horizontal line as the running water surface. "Subtraction" of frames is also accompanied to make it robust to horizontal line like disturbances. A new approach in this report applies "addition" of frames and a "horizontal" edge detector to distinguish water region and land region. Variance of each line of a filtered video frame is used as a feature value. Optimization of the edge detection filter is also discussed so that the system becomes robust to changes of lighting condition.

This paper proposes a new type of layered video coding for an application of river monitoring. A sensor node of the system produces the bit stream functionally separated into three layers. The first layer contains component signals effective for a function of water level detection. The second layer contains signals for thumb-nail video browsing. Each of them is transmitted at very low bit rate for regular monitoring. The third layer contains additional data necessary for decoding original video signal. It is transmitted when the necessity arises. The original video signal is decomposed into component band signals by the Haar transform in a sensor node. Experimental results show which band signals should be included into the first layer for water level detection based on the maximum likelihood estimation. It becomes possible to reduce power consumption in a sensor node by distributing computational load via internet and by sharing the temporal and spatial processing between "compression" and "recognition".

This paper proposes a new type of layered video coding especially for the use of monitoring a river or a water channel. A sensor node of the system decomposes a video signal into some components and produces a bit stream which is functionally separated into three layers. The first layer contains the minimum components effective for detecting water level. The second layer contains signals for thumb-nail video browsing. Each of them is transmitted at very low bit rate for regular monitoring. The third layer contains additional data for decoding the original video signal. It is transmitted in case of necessity. The original video signal is decomposed into band signals as the components by the Haar transform in a sensor node. Experimental results show which band signals should be included into the first layer considering both of water level detection performance and data size to be transmitted.

While conventional DCT outputs real numbers and it is not suitable for lossless coding of images, the integer DCT (Int-DCT) outputs integer numbers and can serve as a lossless coding and lossy coding compatible to the conventional DCT. Recently, our group has proposed an optimum assignment method of word length of multipliers in the Int-DCT assuming "colored" spectrum of input signals to reduce total hardware complexity of the circuit. However, it has been ignored to include variations of the integer rotation transforms. This report theoretically and experimentally investigates various rotation transforms and finds the best combination to construct the Int-DCT least sensitive to finite word length expression of multiplier coefficients. As a result, improvement up to 1.8 [dB] to the existing one was confirmed for a colored signal.

Recently JPEG 2000 has adopted the integer wavelet to attain "lossless" coding of images. However existing DCT based "lossy" coding still prevails all over the world. This paper focuses on the integer DCT which can perform "lossless" coding maintaining compatibility to the "lossy" DCT coding. So far, we have optimized word length of each multiplier in the integer DCT considering colored spectrum of input image signal. However it is remained unknown that 1) Is the assigned word length robust to change of input signal? 2) How does it depend on structure of the transform? This report is intended as an investigation of theoretical analysis to answer such questions.

The lifting structured wavelet transform (lifting wavelet) attracts researchers' attention as a key technology to a lossless and lossy unified coding system of digital image data. For VLSI circuit implementation, multiplier coefficients in the transform must be truncated into finite word length binary values. This report investigates how to appropriately assign error tolerance to each coefficient considering its sensitivity to a typical image signal under a given tolerable total hardware cost. As a result, it is confirmed that the total hardware scale of the multipliers is reduced to 51%.

We propose a new lossless progressive color coding of color still images. An image is firstly transformed by reversible wavelet transform such as S transform due to progressive transmission of image. And we utilize for an improvement of the compression performance not only for the inter color correlation but also the inter/intra band correlation of the subbands of reversible wavelet transform. We show three effective predictors to remove that correlation, and it is confirmed by numerical experiment that combination of all the predictions is the best in performance.

In this report, a digital image data compression method which has two modes-"lossless" mode and "lossy" mode-and also "progressive" functionality is proposed. The coding system can be optimized for an arbitrary input image signal under the lossless/lossy unified coding gain introduced here.

A new lossy/lossless coding of digital image data with two channel non-separable filter bank is proposed. The method has freedom of selecting taps and coefficients of the filters and this freedom is used for maximizing the coding gain. Simulation results are shown to confirm its effectiveness.

In this paper a new lossless coding of digital image data with a four channel filter bank (4ch FB) having optimized filter coefficients is proposed. The method has an advantage of less restrictions on optimizing the filter coefficients compared to the cascade form of two channel filter banks (2ch FB). Simulation results are shown to confirm its effectiveness.

This paper theoretically investigates data distortion caused by resolution conversion in DCT (discrete cosine transform) based layered coding system. The system is attractive because it can not only compress large amounts of video data but also convert the scale of reconstructed pictures. However only a few attempts have so far been made at the degree of aliasing error. In this paper we treat the DCT system as a filter bank so that we can evaluate the system in the frequency domain. As a result of our study the system is theoretically evaluated from the multirate stand point and the DCT system is compared with other layered coding systems, e.g. subband coding and Gaussian/Laplacian pyramid, by means of one objective measure.>

This paper investigates the cause of blurring called drift, which is observed in inter-frame predicted and down scaled pictures with a scalable video coding system, from a multirate stand point of view and proposes a new motion compensation method in the decoder for drift reduction.>

Describes the development of a technique for subband image coding called the symmetric extension method which utilizes the nature of a symmetrically extended image to achieve high quality coding. There are, however, some restrictions on the usable subband coding system. The development is done based on the property of a symmetrically extended signal in the analysis process to remove the restriction. The behavior of the sequence after two processes, filtering and decimation, is examined and formulated. The development extends the application area of the technique.