The abstracts and titles presented here are subject to change based on the temporal evolution of the respective theses.

 

PhD Thesis Students (Ongoing, Updated July 14, 2015)

 

Audio fingerprinting of Film and Classical songs for multi-lingual retrieval

Ph.D. thesis in progress of S. Sri Ranjani (Jan 2010 to date)

 

The core of the project is the development of the sum-of-spectrogram (SOS) feature to capture the emphasis and de-emphasis points in a particular rendition. The collapse of the frequency scale makes the base feature less sensitive to pitch variations (due to singer voice changes male-to-female and vice-versa) and pitch jitters owing to lyrical translations incurred during language changes. While the feature captures the underlying beat or rhythm of the song, it fails to derive sufficient information related to the melodic construct. Its extension as a multi-band SOS, offers some improvement to the algorithm in terms of recovering a part of the melody buried in the song.

 

Future extensions include:

a) Melody extraction and re-synthesis from quantized MULTI-BAND SOS footprints (as proof of verification).

b) Application to RAGA based classification through EXEMPLAR LEARNING.

c) Application to note segmentation in classical renditions, in the absence of any form of accompaniment.

d) Application to GENRE classification (devoid of lyrics and combinations of ragas): GHAZALS, CARNATIC, HINDUSTANI, MULTI-LINGUAL FILM SONGS etc..

e) Probabilistic models associated with raga-based carnatic renditions.

 

Papers published/accepted:

1) S. Sri Ranjani, K. Karthik, P. K. Bora and V. Abdulkareem,”Multi-band sum of Spectrogram based Audio fingerprinting of Indian film songs for Multi-lingual Song Retrieval,” Accepted in International Conference on Advances in Computing, Communications and Informatics (ICACCI 2015), Kochi, Kerala, 2015.

 

2)  S. Sri Ranjani, V. Abdulkareem, K. Karthik and P. K. Bora,”Application of Shazam Based Audio Fingerprinting of Indian Film Songs for Multi-lingual Song Retrieval”, Proceedings of National Workshop on Advances on Communication and Computing (WACC), Guwahati, Assam, India, September 26-27, 2014: published in Springer Lecture notes in Electrical Engineering, Volume 347, pp 81-92, 2015.

 

Forensic analysis of natural still images based on inconsistencies in Geometric, Luminance and Camera aberration  patterns

Ph.D. thesis in progress of Parveen Malik (Aug 2010 to date)

 

The problem began with an intense drive to study the origins, detect and analyze the characteristics of Purple Fringing aberration (PFA) patterns acquired in images captured using low-cost cameras (normal and mobile). Its current extensions include:

a) Synthesis of different forms of PFA patterns in regular images through RGB colour plane rotation.

b) Using directional information (spatial orientations) of PFA patterns and its composition in the Cb-Cr space to quantify the COMPOSITION of the resident LIGHT SOURCEs and their RELATIVE POSITIONs with respect to the scene and camera.

c) Using PFA patterns derived from dark patches and shadow regions to estimate the location of the light source.

 

Other related problems include:

d) Detecting anomalies in GEOMETRIC CUES when images from two different perspectives are SPLICED.

e) Using OBJECT IMMERSIONS to detect forgeries in natural images (natural objects are inpainted with the background).

 

Papers published/accepted:

1) P. Malik and K. Karthik,”Detection of Purple Fringing Aberration Patterns and its Application to Forensic Analysis of Spliced Images”, Accepted in International Conference on Advances in Computing, Communications and Informatics (ICACCI 2015), Kochi, Kerala, 2015.

 

Liveliness detection without any prior training pertaining to prosthetics

Ph.D. thesis in progress of Balaji Rao (Dec 2012 to date)

 

Given a stored sequence of still images either comprising of pose variations pertaining to a particular subject or a face-video sequence of a living individual, is it possible to detect the onset of prosthetics without any prior training?

 

Future extensions include:

a) Registering a set of poses associated with a particular subject: This problem has multiple levels of complexity:

a-1) Registration when there is only rotation (left/right or up/down or diagonal movements) - camera, light source and the centroid of the individual's face are fixed in space.

a-2) Registration when there are illumination variations associated with each pose not just because of difference in the orientation of the light source but also due to its differences in spectral composition.

a-3) Registration of poses when partial disguises or natural cosmetic variations are involved such as beards, changes in hair-style, ageing etc...

a-4) Studying the importance of the location of the light source from the point of view of detecting liveliness in faces.

 

[Applications in BIOMETRIC AUTHENTICATION and SPOOF detection].

 

b) Isolating caricatures from a set of natural poses mixed with prosthetics/morphed variations of a specific face. 

 

[Applications in forensic analysis and detection of spliced and morphed face variations of celebrities posted in social networks].

 

Other problems include:

c) Automated lip-reading from surveillance video sequences.

d) Emotion recognition based on the EYES alone and the EYES and NOSE of a particular subject.