Shankar Mosur Venkatesan

Dynamic illumination based system to remove the glare and improve the quality of medical images

Medical images taken from camera based devices (e.g. laparoscope, colposcope, retinoscope, etc) are greatly affected by numerous bright reflection spots (called glare or specular reflections). This may affect the visibility of the abnormal features (if present in the glare locations). We have developed a novel solution to overcome this problem by incorporating a multi-LED lighting solution. This will intelligently and rapidly switch on and off the LEDs in a pattern that dynamically and geometrically shifts/shuffles these glare spots back and forth in the image such that every glare-affected area of a single image frame can be reconstructed from a few adjacent time-frame images. We have built the prototype that successfully demonstrates how the glare problem in the medical video/image can be satisfactorily solved, significantly enhancing the accuracy of this vital procedure in the diagnosis of diseases. We achieve 65-95% reduction in specularity on phantom model using the proposed approach.

Tensor Voting based Lazy Learner for Pattern Classification

In this paper, we propose a lazy learning classifier based on tensor voting framework for supervised binary and multiclass problems. Unlike other lazy learners, our approach communicates votes as tensors which allow them to communicate more information about the local structure/orientation. Hence, classification of a new datapoint is not only based on its proximity to training datapoints but also its structural alignment. The only variable parameter in our approach is the scale of voting. Our experiments on benchmark datasets demonstrate the efficacy of the proposed approach.

Kernel ridge regression for supervised classification using Tensor Voting

In this paper, we propose classifiers based on Tensor Voting (TV) framework for supervised binary and multiclass problems. Traditional classification approaches classify a test sample or point based on its proximity to classes of a training set, where proximity is generally taken as some variant of the Euclidean distance in the original or some transformed higher dimensional space. However, we may need to have more intrinsic and computable features of samples than just Euclidean position since classes or patterns usually live on generalized manifolds (that might change in time as well) and thus need an easily parametrizable and computable but powerful notion of “distance” for classification of test samples that may be closer to a class manifolds by distance as well as by orientation or curvature. Our classification approach takes a step in this direction and infers a “local orientation or structure” by computing an already known tensor representation of the training set and performs classification based on the usual distance as well as the estimated (or given) local orientation. In our paper, we describe a novel Eager classification scheme where the central idea is that a kernel ridge regression is carried out on the TV output during the training phase itself so that only evaluation using the kernel is needed during the test phase, achieving a significant speed-up in testing time on benchmark datasets when compared to the lazy learning classifier without much compromise in classification accuracy. The only variable parameter in our approach is the scale of the voting. Our experiments on benchmark datasets demonstrate the efficacy of the proposed approach.