DSCN9426We mainly working in the confluence of computer vision and machine learning.

Objectives:

  • exploring features to describe images, audio and video content
  • automatic segmentation into meaningful units
  • classification into genres
  • exploring compressive sensing, deep learning and kernel methods for the above

Issues in computer vision (video content analaysis):

  • choice of the features that are compact, robust to illumination changes, camera/object motion and descriptive of the content.
  • detection of temporal discontinuities in video sequences
  • detection of continuous sequence of shots having a common semantic significance
  • categorizing a given video into one of the predefined classes or genre(s) with the information from above

Techniques explored:

  • Convolutional Neural Networks
  • Deep Neural Networks
  • Hidden Markov Models (HMMs)
  • Sparse coding & dictionary learning
  • Auto-associative neural networks

 
Research Impact:

  • The important issue in the large-scale de-duplication applications is the speed and accuracy of the matching as the number of persons to be enrolled runs into millions. Techniques for de-noising and classification of biometrics are proposed to speed up the de-duplication process. Various methods related to GPGPU (General Purpose Graphical Processing Unit) processing will be proposed to parallelize the different phases in biometric recognition system.
  • Content based medical image classification and retrieval using sparse representation and dictionary learning techniques were proposed for both labeled and unlabeled data. An online dictionary learning technique is proposed for modality based medical image classification (MRI, MRA,DTI and FMRI). An ECG analyzer to classify the heartbeat of patients suffering with Cardiovascular Diseases (CVD) into normal and abnormal is developed using various kernel based classification techniques.