Amir Najmi

Image classification by a two-dimensional hidden Markov model

For block-based classification, an image is divided into blocks, and a feature vector is formed for each block by grouping statistics extracted from the block. Conventional block-based classification algorithms decide the class of a block by examining only the feature vector of this block and ignoring context information. In order to improve classification by context, an algorithm is proposed that models images by two dimensional (2-D) hidden Markov models (HMMs). The HMM considers feature vectors statistically dependent through an underlying state process assumed to be a Markov mesh, which has transition probabilities conditioned on the states of neighboring blocks from both horizontal and vertical directions. Thus, the dependency in two dimensions is reflected simultaneously. The HMM parameters are estimated by the EM algorithm. To classify an image, the classes with maximum a posteriori probability are searched jointly for all the blocks. Applications of the HMM algorithm to document and aerial image segmentation show that the algorithm outperforms CART/sup TM/, LVQ, and Bayes VQ.

Image classification by a two dimensional hidden Markov model

Traditional block-based image classification algorithms, such as CART and VQ based classification, ignore the statistical dependency among image blocks. Consequently, these algorithms often suffer from over-localization. In order to benefit from the inter-block dependency, an image classification algorithm based on a hidden Markov model (HMM) is developed. An HMM for image classification, a two dimensional extension from the one dimensional HMM used for speech recognition, has transition probabilities conditioned on the states of neighboring blocks from both directions. Thus, the dependency in two dimensions can be reflected simultaneously. The HMM parameters are estimated by the EM algorithm. A two dimensional version of the Viterbi algorithm is also developed to classify optimally an image based on the trained HMM. An application of the HMM algorithm to document image and aerial image segmentation shows that the algorithm performs better than CART.

An interactive dialog system for learning Japanese

Subarashii is a system that uses automatic speech recognition (ASR) to offer first-level, computer-based exercises in the Japanese language for beginning high school students. Building the Subarashii system has identified strengths and limitations of ASR technology. The system was tested with 34 students at Silver Creek High School in San Jose, California and with 13 students at Stanford University in Stanford, California. Recognition accuracy was measured and user errors were analyzed. The functional accuracy defined as the percentage of time when the system performs the correct functional behavior turned out to be generally higher than the per-utterance speech recognition accuracy.

A Potential Biomarker in the Cord Blood of Preterm Infants Who Develop Retinopathy of Prematurity

Preterm infants are at risk of developing sepsis, necrotizing enterocolitis (NEC), chronic lung disease (CLD), and retinopathy of prematurity (ROP). We used high-throughput mass spectrometry to investigate whether cord blood proteins can be used to predict development of these morbidities. Cord blood plasma from 44 infants with a birth weight of <1500 g was analyzed by surface-enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI-TOF). Six infants developed ROP ≥stage II, 10 CLD, three sepsis, and one NEC. We detected 814 protein signals representing 330 distinct protein species. Nineteen biomarkers were associated with development of ≥stage II ROP [false-discovery rate (FDR) <5%] and none with CLD. Several proteins with molecular weight (Mr) 15–16 kD and pI 4–5 were detected with increased abundance in infants with ROP, while similar Mr proteins with pI 7–9 were less abundant in these patients. Sodium dodecylsulfate–polyacrylamide gel electrophoresis (SDS-PAGE) and sequence analysis identified these proteins as α-, β-, and γ-globin chains. Partial deamidation of Asn139 in β-globin chains was observed only in the pI 4–5 proteins. We conclude that there are several promising biomarkers for the risk of ROP. Deamidation of globin chains is especially promising and may indicate underlying prenatal pathologic mechanisms in ROP. Validation studies will be undertaken to determine their clinical utility.

Speech recognition in a system for teaching Japanese

Automatic speech recognition (ASR) is implemented and tested in a system to offer first‐level, computer‐based exercises in the Japanese language for beginning high school students. The paper describes a development project designed to identify the strengths and limitations of current ASR technology in the development of materials for computer‐assisted interactive spoken language education. With current ASR technology, several kinds of pedagogically useful exercises can be constructed that should aid spoken language acquisition in a non‐native student. A system design is presented, along with some example exercises and a description of methods that are used to evaluate its effectiveness with students at different levels of proficiency. At the present time, there is no agreed‐upon set of ASR functions that is adequate to support many common types of teacher–student interactions. The project has begun to address such issues in the course of developing and field testing a representative set of exercises. [Wor...

Blind denoising using a wavelet coder

We present a method based upon data compression, for denoising an image corrupted by white Gaussian noise. Our work differs from that of others not only in its statistical/information theoretic basis but also in that our approach extends to denoising problems where the noise variance is unknown.

A criterion for model selection using minimum description length

Rissanen (1978) proposed the idea that the goodness of fit of a parametric model of the probability density of a random variable could be thought of as an information coding problem. He argued that the best model was that which was able to describe the training data together with the model parameters using the fewest number of bits of information (Occams razor). This paper builds upon that basic insight and derives a more general result than did Rissanen, dealing as he was, with time series analysis. To arrive at a model selection criterion with wider applicability, the present derivation relies upon results from information theory and the theory of rate-distortion.