Arijit Laha

Design of vector quantizer for image compression using self-organizing feature map and surface fitting

We propose a new scheme of designing a vector quantizer for image compression. First, a set of codevectors is generated using the self-organizing feature map algorithm. Then, the set of blocks associated with each code vector is modeled by a cubic surface for better perceptual fidelity of the reconstructed images. Mean-removed vectors from a set of training images is used for the construction of a generic codebook. Further, Huffman coding of the indices generated by the encoder and the difference-coded mean values of the blocks are used to achieve better compression ratio. We proposed two indices for quantitative assessment of the psychovisual quality (blocking effect) of the reconstructed image. Our experiments on several training and test images demonstrate that the proposed scheme can produce reconstructed images of good quality while achieving compression at low bit rates.

Land cover classification using fuzzy rules and aggregation of contextual information through evidence theory

Land cover classification using multispectral satellite images is a very challenging task with numerous practical applications. We propose a multistage classifier that involves fuzzy rule extraction from the training data and then the generation of a possibilistic label vector for each pixel using the fuzzy rule base. To exploit the spatial correlation of land cover types, we propose four different information aggregation methods which use the possibilistic class label of a pixel and those of its eight spatial neighbors for making the final classification decision. Three of the aggregation methods use the Dempster-Shafer theory of evidence, while the remaining one is modeled after the fuzzy k-NN rule. The proposed methods are tested with two benchmark seven-channel satellite images, and the results are found to be quite satisfactory. They are also compared with a Markov random field model-based contextual classification method and found to perform consistently better.

An unbalanced data classification model using hybrid sampling technique for fraud detection

Detecting fraud is a challenging task as fraud coexists with the latest in technology. The problem to detect the fraud is that the dataset is unbalanced where non-fraudulent class heavily dominates the fraudulent class. In this work, we considered the fraud detection problem as unbalanced data classification problem and proposed a model based on hybrid sampling technique, which is a combination of random under-sampling and over-sampling using SMOTE. Here, SMOTE is used to widen the data region corresponding to minority samples and random under-sampling of majority class is used for balancing the class distribution. The value difference metric (VDM) is used as distance measure while doing SMOTE. We conducted the experiments with classifiers namely k-NN, Radial Basis Function networks, C4.5 and Naive Bayes with varied levels of SMOTE on insurance fraud dataset. For evaluating the learned classifiers, we have chosen fraud catching rate, nonfraud catching rate in addition to overall accuracy of the classifier as performance measures. Results indicate that our approach produces high predictions against fraud and non-fraud classes.

Fast codebook searching in a SOM-based vector quantizer for image compression

We propose a novel method for fast codebook searching in self-organizing map (SOM)-generated codebooks. This method performs a non-exhaustive search of the codebook to find a good match for an input vector. While performing an exhaustive search in a large codebook with high dimensional vectors, the encoder faces a significant computational barrier. Due to its topology preservation property, SOM holds a good promise of being utilized for fast codebook searching. This aspect of SOM remained largely unexploited till date. In this paper we first develop two separate strategies for fast codebook searching by exploiting the properties of SOM and then combine these strategies to develop the proposed method for improved overall performance. Though the method is general enough to be applied for any kind of signal domain, in the present paper we demonstrate its efficacy with spatial vector quantization of gray-scale images.

Accelerated codebook searching in a SOM-based Vector Quantizer

Kohonens SOM algorithm has been used successfully by some researchers for designing codebooks. However, while performing an exhaustive search in a large codebook with high dimensional vectors, the encoder faces a significant computational barrier. Due to its topology preservation property, SOM holds a good promise of being utilized for fast codebook searching. In this paper we develop a method for fast codebook searching by exploiting the topology preservation property of SOM. This method performs non-exhaustive search of the codebook to find a good match for a input vector. The method is a general one that can be applied to various signal domains. In the present paper its efficacy is demonstrated with spatial vector quantization of gray-scale images.

Extraction of contextual information from medical case research report using WordNet

Relevant information within a document are usually embedded within a few sentences or passages (units). If any semantic tagging can be associated at the unit level within a document, the understanding of the information will be deeper and quicker saving a lot of effort and time of the user. In this paper we propose a simple approach of sentence tagging using the relational semantic network among lexical units as presented in WordNet. The approach is to propose a domain specific sub-taxonomy of key concepts following WordNet structure and associate a meaning with each of the sentences contextually. This approach also identifies those words from the text that can provide important semantic information in a tag assignation task. The occurrence of keywords will determinate a series of patterns that can be converted into rules for deciding the tagging and also information extraction as a useful application.

Compositional information extraction methodology from medical reports

Currently health care industry is undergoing a huge expansion in different aspects. Advances in Clinical Informatics (CI) are an important part of this expansion process. One of the goals of CI is to apply Information Technology for better patient care service provision through two major applications namely electronic health care data management and information extraction from medical documents. In this paper we focus on the second application. For better management and fruitful use of information, it is necessary to contextually segregate important/ relevant information buried in a huge corpus of unstructured texts. Hence Information Extraction (IE) from unstructured texts becomes a key technology in CI that deals with different sub-topics like extraction of biomedical entity and relations, passage/paragraph level information extraction, ontological study of diseases and treatments, summarization and topic identification etc. Though literature is promising for different IE tasks for individual topics, availability of an integrated approach for contextually relevant IE from medical documents is not apparent enough. To this end, we propose a compositional approach using integration of contextually (domain specific) constructed IE modules to improve knowledge support for patient care activity. The input to this composite system is free format medical case reports containing stage wise information corresponding to the evolution path of a patient care activity. The output is a compilation of various types of extracted information organized under different tags like past medical history, sign/symptoms, test and test results, diseases, treatment and follow up. The outcome is aimed to help the health care professionals in exploring a large corpus of medical case-studies and selecting only relevant component level information according to need/interest.

An Empirical Study on the Robustness of SOM in Preserving Topology with Respect to Link Density

Practical implementations of SOM model require parallel and synchronous operation of the network during each iteration in the training stage. However this implicitly implies existence of some communication link between the winner neuron and all other neurons so that update can be induced to the neighboring neurons. In the current paper we report the results of an empirical study on the retention of topology preservation property of the SOM when such links become partially absent, so that during a training iteration not all the neighbors of the winner may be updated. We quantify our results using three different indexes for topology preservation.