I have another set of interests (or vices) in Number Theory, Cryptanalysis and Programming.
Apart from these professional interests, I am also an amateur astronomer, model railroad enthusiast (now you know what happens during vacations!), cruciverbalist (go look it up on Google or whatever!) and ... oh yes, intermediate level player, armchair critic and follower of many sports including cricket, tennis, badminton and golf.
There is also a need somewhere to fit in PGW, Agatha Christie, Asterix, Sherlock Holmes, Shaun the Sheep, Mr. Bean, Looney Tunes, etc. Finally, I need to account for time spent on Carnatic music and Classic Rock - days sure fly by swiftly.
Telugu (తెలుగు) Optical Character
Recognition:
An Optical Character Recognition (OCR) system converts document
images into text (either in ASCII or Unicode formats). OCR systems
generally have a preprocessing stage for cleaning input document
images and segmenting them into lines, words and characters while
separating out graphical and other non-text elements. The text
elements are passed to a classifier which labels them for output in
text format. Our classifier is a 5-layer convolutional neural
network trained on about 70,000 labelled and 200,000 unlabelled
samples.
We have a complete end-to-end OCR system for Telugu script in our
lab. It has been tested on a 4000 page third-party corpus and gives
an accuracy greater than 91%. Most of the errors occur due to
preprocessing, especially incorrect segmentation of characters. When the
characters are correctly segmented, the accuracy is greater than 96%.
Colour Image Processing:
Colour images are not, repeat not three times grayscale
images. Colour is much more interesting and fascinating with strong
links between physics, human vision system and digital spaces and
models.
In our lab, we do research on colour science exploring links between spectra and colour operations, developing new colour operations with strong links to physical processes, use of colour in image forensics, and use of human vision principles to improve performance of deep networks.
There is also a Masters level course on Colour Image Processing offered as an elective during January-May semester.
Demosaicing (right) an image from a digital camera (left)
Deep Learning:
If you are a Pattern Recognition or an Image Processing guy, well,
what else can you do these days? Like it or not, deep learning is
the rage and I am not immune to it.
My interests lie in using deep learning to study cognitive processes (at least, informally), work on theories behind deep learning by looking at stochastic processes and, finally, in applying deep networks to various day-to-day tasks and have some fun!
We have several students working on deep networks in OCR, video processing, geometry and computer vision tasks, k-shot, one-shot and zero-shot learning, etc. Input domains are primarily visual and text with a smattering of speech.
Hooked to astronomy from the time I was eight years old or so. Astronomy gave me one of my happiest moments as also one of the greatest disappointments. I saw an occultation of Saturn by the Moon in the year 2000 (I think!): an absolutely breath-taking and awe-inspiring sight to see 'tiny' Saturn emerge from behind a 'giant' Moon. Talk about a live and natural lesson in perspective and 3-D geometry! I cannot forget the sight - it is too emotional, even now gives goosebumps to think about it.
I was in Intermediate first year when the biggest disappointment (in hindsight!) of missing a total solar eclipse happened; on 16 February 1980. The eclipse was 99.5% in Vidyanagar, Hyderabad and was total from any place about 4 km south, and I didn't go - how was I to know that a Solar Eclipse is the grandest celestial spectacle and that the next one in Hyderabad is scheduled only for 18 September 2248! Although, at that time, I must admit, even 99.5% was was quite spectacular - we even had shadow bands scurry across the ground and could see the darkened cone race across the south-eastern sky.
Click on any picture to get a full resolution version.
... is hard! You need the right equipment: a solid tripod, a good DSLR with settings for time or bulb exposures and a motor drive to track the sky.
My equipment: a 12-yr old Canon point-and-shoot camera, a Celestron 8-inch and a not-so-good tripod. Add light-pollution in Hyderabad and things become quite tricky, to say the least.
All photos here use afocal method where the camera shoots through the eyepiece: usually 12.5mm or 25mm. Exposures are less than 10s.
The most recognisable constellation of all! Visible from almost anywhere on the Earth, it is also a pointer to many other constellations for a beginner.
Taken about 2 days before the full moon, it shows almost all the principal features on the Moon. The bright crater at the bottom is Kepler and the small bright spot near the left-centre is Aristarchus: one of the brightest regions on the Moon.
Mare Imbrium (Sea of Showers) region with the crater Copernicus on the left. One of my favourites on the Moon because of the variety of features. The almost circular mountain range is punctuated by the prominent crater Plato (on the right).
Astronomy provides a wonderful field to enjoy rarities. Some of
these are eclipses, transits, occultations and conjunctions.
Partial Solar Eclipse
15 January 2010
For one day, 6 January 2018, Jupiter had a new satellite: Mars! The red planet was only 0.1 degree from Jupiter. In this photo, the bump on Jupiter is Io, then Europa and Ganymede and the farthest, Callisto.
Save the file spiro-designer. If Tcl/Tk is properly installed, you can run it by typing either wish -f spiro-designer or ./spiro-designer.
Poster Presentations:
The item on the right is a portal into a Virtual Reality room displaying posters about research done by some faculty of the school. The page is based on the Aframe technology. You can navigate the room as follows: