Welcome to my home page !
MS, PhD Theses | Resume | Citations | Patent | Peer-Review | GitHub | LinkedIn
Click on each project title to view its description.
Note: I am the 1st author in all project publications listed below.
InSAR Image-based Wide Area Monitoring via Deep / Machine Learning & Computer Vision
Wide Area Monitoring System: Combining Machine Learning, Deep Learning & Computer Vision techniques to enhance the accuracy and speed of large scale motion mapping of ground displacement caused by groundwater extraction, mining, oil & gas, urban development etc. from Inteferometric Synthetic Aperture Radar (InSAR) satellite images. Developed first ever Convolutional Neural Network-based filtering and point-wise signal quality quantification methodology for InSAR, and further improved it to CNN-guided generative modeling-based approach.
Implemented in Keras with Tensorflow-GPU back-end and Python, supported by the numpy, scipy, scikit-image, scikit-learn and matplotlib libraries in an Anaconda environment.
“CNN-based InSAR Coherence Classification”, 17th IEEE Sensors, Oct 28-31, 2018, New Delhi, India (preprint | published | code)
“CNN-based InSAR Denoising and Coherence Metric”, 17th IEEE Sensors, Oct 28-31, 2018, New Delhi, India (preprint | published | code)
“An Unsupervised Generative Neural Approach for InSAR Phase Filtering and Coherence Estimation”, IEEE Geoscience and Remote Sensing Letters (preprint | published | code)
Image Quality from Deep Learning ( Opinion-&-Distortion-Unaware, No-reference IQA )
Algorithm-based Image Quality Assessment outputs a quality score for a given (possibly distorted) input image to mimic the response of a human observer. Traditional IQA required a distortion-free version of the input image (full-reference), knowledge of types of possible distortions (distortion-aware) or training on subjective opinion scores (opinion-aware) and was based on hand-crafted features. This project proposes and validates the first ever method to overcome all of these limitations using learned features.
“Potential of deep features for opinion-unaware, distortion-unaware, no-reference image quality assessment”, International Conference on Smart Multimedia (Springer), Dec 16-18, 2019, San Diego, USA (preprint | published | code)
Image-based parameter-tuning of a noise filter using Deep Learning
This project explores a new direction in explainable Artificial Intelligence (AI). Most state-of-art deep-learning based noise removal algorithms involve some form of neural networks trained end-to-end (input: noisy image, output: clean image). Such methods are inexplicable "black-boxes" since we do not have a clear understanding of their inner workings and how change in input data distribution can affect their denoising performance. In contrast, this project proposes and validates a CNN-based method that tunes the parameter of a computer vision based (fully explainable) denoising algorithm based only on the noisy input image. Thus, we get the best of both worlds: explain-ability of vision and learning-based regression.
“CNN-based Real-Time Parameter Tuning for Optimizing Denoising Filter Performance”, 16th ICIAR (Springer), 27th to 29th August 2019, University of Waterloo, Canada (preprint | published | code)
White Matter Injury detection in pre-term infants' T1 MR Images using Computer Vision
Injury to the white matter and periventricular regions in infant brains may serve as early predictors of developmental deficits. This project detects WMI from T1 MR images of pre-term infants, which is specially challenging due to lack of brain atlas, small size of the brain, short scan duration and their constant movement during scanning. This creates very low-resolution and extremely noisy MR images, making this project challenging. I developed the first fully automated WMI detection method that does not require brain atlas and heuristically approximates tissue segmentation, greatly reducing computation.
“A Fast Segmentation-free Fully Automated Approach to White Matter Injury Detection in Preterm Infants”, Medical and Biological Engg. & Computing (SCI Indexed, Springer, Impact Factor: 2.04) Vol 57, Issue 1, pp 71-87 (preprint | published | code)
Mobile GPU-friendly coding artifact reduction in High Dynamic Range images
Dolby Vision codecs encapsulate traditional ones like AVC/HEVC, and provide HDR capability. However, banding artefacts need to be handed while displaying SDR content on HDR screens. Computationally constrained mobile GPU environments restrict use of traditional filtering-based methods. I developed a dithering-based solution. It operates on individual pixels and modulates noise injection based on slope of Inverse Tone Mapping curve. Use cases also include those where the input is quantized due to bit-depth conversion, but the unquantized original version is unavailable. In another project, I approximated the tone-mapping curve (via interpolation) using fewer number of points without any perceptible quality degradation. This approximation resulted in a four-fold improvement in performance of the respective HDR processing pipeline. I conducted subjective experiments on the Dolby Pulsar professional reference monitor to validate my proposed methods.
Simulations/prototyping done in Matlab and implementations in the C programming language.
“Adaptive Dithering using Curved Markov-Gaussian Noise in the Quantized Domain for Mapping SDR to HDR Image”, International Conference on Smart Multimedia (Springer), Aug 24-26, 2018, Toulon, France (preprint | published)
Versatile Depth-of-Field effects and Image Auto-ROI
This is somewhat similar in purpose to the "Lens Blur" feature in the Google Camera android app, though the methodology and input modality are completely different:
1. Formulation of a novel stereo disparity estimation algorithm based on sparse disparity estimates. Here, only disparities of segment boundaries of the left image are estimated, which form the basis for interpolating the remaining disparities. Thus, the stereo matching process becomes time-efficient.
2. Part of the computationally intensive tasks are offloaded to GPUs using Java-APARAPI, while the rest are dispatched to the multiple cores of the CPU using the Java Thread Pool (JTP). This results in parallel processing and reduced execution time.
3. Using the depth map output by the proposed algorithm, selective Gaussian blurring of 3D video scenes based on the depths of non-interest to the user were carried out, simulating artificial depth-of-field effects found in professional cameras. This is useful in portraiture, macro-photography etc.
4. Depth information acquired from sensor modalities like stereo is used along with saliency to auto-emphasize objects of interest in a scene, thus increasing its visual appeal without user intervention.
5. Design & validation of novel entropy-based confidence measure for stereo error detection.
“Highlighting Objects of Interest in an Image by Integrating Saliency and Depth”, 23rd IEEE International Conference on Image Processing (ICIP 2016), Phoenix, USA, 25th to 28th September 2016 (preprint | published | code)
“Entropy-difference based Stereo Error Detection”, 12th IEEE Image Video and Multidimensional Signal Processing (IVMSP 2016) workshop, Bordeaux, France, 11th and 12th July 2016 (preprint | published)
“Depth-based Selective Blurring in Stereo Images Using Accelerated Framework”, 3D Research (ESCI Indexed, Springer, CiteScore: 1.02) Vol 5, Issue 3, September 2014 (preprint | published | code)
“A Hybrid Algorithm for Disparity Calculation from Sparse Disparity Estimates Based on Stereo Vision”, 10th IEEE International Conference on Signal Processing and Communications (SPCOM), Indian Institute of Science (IISc), Bangalore, India, 22nd to 25th July 2014 (preprint | published | code)