Werapon Chiracharit

Exudates detection in fundus image using non-uniform illumination background subtraction

Exudates detection in fundus image using non-uniform illumination background subtraction is proposed in this paper. Non-uniform illumination background that is the major problem of fundus imagge processing is need to be eliminated. Weighted surface fitting is used in fundus image background estimation after image compensation is performed. The unwanted data for fundus image background estimation Le. optic disk, fovea, blood vessels, and lesions are detected without traditional image segmentation. The estimated background is subtracted from the image and then exudates are detected from the foreground of image using level-set evolution without re-initialization. The experimental results show that the proposed method is robust to non-uniform illumination environment. Averaging sensitivity and positive predictive value are 91.84 % and 92.81 % respectively.

An architecture of HMM-based isolated-word speech recognition with tone detection function

In an architecture of speech recognition for some languages (such as Thai, Chinese, and so on) that tone plays a key role in meaning classification, tone detection function is required in order to guarantee a correct word recognition. This paper proposes an architecture of HMM-based isolated-word speech recognition with tone detection function. In this architecture, tone detection function is added into each computation process of series architecture and each parallel computation of scalable architecture. To evaluate the performance of the proposed method, the experiment is set and performed with 29 Thai words selected from TV remote control commands and 10 Thai indistinct tone classification words. The results reveal 4.94% improvement of accuracy rate for remote control commands and 10.75% for indistinct tone classification words comparing with the conventional architecture.

Reduced complexity tone classifier for automatic tonal speech recognizer

A tone classifier is an essential part of an automatic tonal speech recognizer (ATSR) because tonal languages recognize word meaning by tones. However, many researchers have developed a highly efficient tone recognition by using rich mathematical techniques and used the whole input speech as an input of pitch detection process. This paper proposes a reduced complexity tone classifier for the automatic tonal speech recognizer. The classifier reduces the number of input frames by detecting only the vowel signals as an input of the pitch detection, called vowel-AMDF (V-AMDF). The classifier uses a lower number of floating-point operations (FLOPs) than used in the whole input speech method. Due to the reduced number of FLOPs, this tone classifier can be suitable for portable electronic equipment. In addition, V-AMDF reduces F0 contour errors caused by the influence from neighboring syllables. This proposed classifier was tested and set by 19 Thai words, selected from voice activation for GPS system and phone dialing options. The experimental results show 86.0% recognition accuracy, and 21.8% reduction in the number of FLOPs, compared with using the whole input speech.

Tone model enhancement for low complexity tone recognition

This paper proposes tone model enhancement for low complexity tone recognition. The tone model reduces the number of input frames by estimating fundamental frequency (F0) from only estimated vowel signals, called vowel magnitude difference function, vowel-MDF (VMDF). Accordingly, it reduces F0 negative influence from neighboring syllables in continuous speech. We enhance tone recognition accuracy by more precise and low calculation vowel segmentation. This results in low complexity tone recognition making it suitable for portable equipment. In addition, the tone model is designed with parallel processing which results in real-time processing. The proposed method was tested and set by 40 Thai words, selected from voice activation for GPS systems and phone dialing options. The experimental results show 8.2% improvement in recognition accuracy, compared with the fixed vowel threshold. The tone recognition provides a 33.8% reduction in number of frames and 25.0% reduction in processing time, compared with using the entire syllable method.

A VLSI architecture of color model-based tonsillitis detection

Tonsillitis disease is the cause of heart attack and pneumonia. It is also a sign of suspected symptom of heart disease. To improve data transfer rates, this paper proposes VLSI architecture by using color model for early-state tonsillitis detection. In this method, Input image is divided into 9 blocks. Each block has 3x3 window which send color data and pixel address to computation box. The system compares the feature values between the normal tonsillitis disease image and the image under detection. These feature values would be input into system for decision support system. Prototype architecture is designed to implement in tonsillitis detection. It has been modeled in VHDL.

Image interpolation by Estimation and Deconvolution of Wavelet Approximate Subband

This paper proposes a method of image interpolation by estimation and deconvolution of wavelet approximate subband. In this method, wavelet approximate subband that contains much information is firstly interpolated by interpolation methods such as cubic spline in order to find the decimated low-pass component of the image. The low-pass component is then deconvoluted for getting the original image. To confirm the effectiveness of proposed method, the experimental results performed with standard test images reveal that proposed method outperforms conventional methods. The best result in peak-signal-to-noise-ratio is 39.27 dB by our proposal with cubic spline interpolation.

Automatic lung segmentation in chest radiographs using shadow filter and local thresholding

Lung segmentation is one of the essential steps in order to develop a Computer-aided Diagnosis (CAD) system for detection of some chest diseases in chest radiographs such as tuberculosis, lung cancer, atelectasis, etc. This paper proposes an unsupervised learning method for lung segmentation in chest radiographs based on shadow filter and local thresholding. The approach consists of three processes: pre-processing, initial lung field estimation and noise elimination. For the first step, the original images are resized and contrast enhanced. Then, each lung outlines are enhanced by shadow filter. The initial lung field estimation are obtained based on local thresholding, delete outer body regions, fill holes and filter regions from their property. However, noise has occurred in the result. To eliminate the noise, morphological operations techniques are used. To evaluate the performance, the proposed method was tested on a public JSRT dataset of 247 chest radiographs. The performance measures of proposed method (overlap, accuracy, sensitivity, specificity, precision, and F-score) are above 90%. The accuracy and overlap are 96.95% and 90.32% respectively with the average execution time of 18.68 s for 512 by 512 pixels resolutions. According to experimental results, our proposed method is unsupervised learning method, no training required and performed accurately.

Improved sub-block encoder of irregular LDPC code for image transmission

This paper presents irregular low-density parity-check (LDPC) code for image transmission system with any desired code lengths. Parity position of sub-block encoder is easily constructed over additive white Gaussian noise (AWGN) channel. The code is used as forward error correction (FEC) technique to protect transmitted source data over the channel. The sub-block encoder is designed with various AWGNs at three code lengths of 522, 2,021 and 4,183. The proposed irregular LDPC code reduces bit error when the code length increases. The performance is evaluated by bit error rate (BER) improvement 33.33% and peak signal-to-noise ratio (PSNR) more than 30 at SNR equals 7 dB.

Scalable architecture of tone classification function for tonal speech recognizer

Tone classification function is used for improving recognition accuracy in tonal speech recognizer (TONE-SPEC). Although average magnitude difference function (AMDF) is generally used to find pitch period of fundamental frequency, there are many frame-repeated processes. This paper proposes scalable architecture of tone classification function for tonal speech recognizer. In the proposed architecture, the number of frames is reduced using vowel-AMDF (V-AMDF). Moreover, there is no frame iteration because the architecture converts series computation of conventional tone classification function into parallel. The parallel computation is designed to be able to reduce or extend the number of frame. Our architecture is set and evaluated with 10 Thai words selected from TV remote control commands and the words having the same phoneme but different tones. The experimental results show that the time consuming of general AMDF and series V-AMDF are improved 85.2% and 72.7%, respectively.

Clustered microcalcification classification using CC-MLO-View corresponding shape and distribution features

Shape of single microcalcifications (muCa++s) and distribution of them in a cluster are two key features for a radiologist to diagnose this abnormality appearing on mammograms into benign type or malignant type of breast cancer. These two features from two-dimensional (2-D) mammogram image from two mammographic views, cranio-caudad view (CC) and medio-lateral oblique view (MLO), are inevitable conflicted because of lack of depth information. It makes a large contradictory information of the same microcalcification cluster in different view. This paper proposes to use three-dimensional (3-D) shape and distribution features exacted from the view correspondence. To identify a 3-D position of microcalcifications, the candidate pairs in CC view and MLO view are stereo-matched based on their relative intensity and size. Occluded microcalcifications are separated by x-ray absorption property. The 3-D shape features are represented by their structural outline, spherical measurement, and thickness which are computed from Fourier descriptor of surface outline, compactness and its intensity, respectively. The distribution feature is represented by 3-D cluster size, average distance between each microcalcifications, and cluster density. There are 12 features used as input features for three-layer feed-forward backpropagation neural network classifier which is constructed dynamically and weighted be training with forty benign and forty malignant microcalcifications. The evaluated performance of the proposed method is 96 percent sensitivity and 91 percent specificity.