Research Projects
Self-similarity in Fibonacci gratings
Fractal signatures are present in the spectrum of Fibonacci gratings because they are quasi-periodic. This gives rise to self-similarity. We showed that the grating structure can be reconstructed using a small sub-band (without the central frequency) information when combined with sparsity. The figure shows a comparison of reconstruction with and without our method. This work was done at the Indian Institute of Technology, Delhi with Dr. Kedar Khare.
Super-resolution using sparsity in Gerchberg-papoulis extrapolation
Modification of the classical Gerchberg-papoulis (GP) method was done by exploiting sparsity. We showed the reconstruction of an object from the knowledge of a limited section of its Fourier spectrum using total variation as a space domain constraint. There was a noted improvement in resolution beyond the diffraction limit. We also found more robust reconstructions from noisy data. The L2-norm reconstruction error significantly went down after combining GP with sparsity. This work was done at the Indian Institute of Technology, Delhi with Dr. Kedar Khare.
Resolution enhancement using Fourier ptychographic microscopy
Proof of concept for Fourier ptychographic microscopy was implemented. We generated multiple low- resolution images using different pass-bands in the Fourier space to mimic structured illumination. Then these images were combined together in the Fourier space using iterative phase retrieval. The figure shows the Fourier space pass-band being doubled and hence improving resolution by two times. We also fabricated a LED array to mount on top of an inverted light microscope and acquired images with different spatial frequency information. This was my final year undergraduate project in collaboration with Tata Institute of Fundamental Research (TIFR), Mumbai.
High-resolution holographic microscope for 3-D imaging of live biological cells
The digital holographic microscope (DHM) recovers the lost phase using computational holography. We developed a tool for image segmentation of cervical cancer cell images. The segmented area was then combined with the phase information to reconstruct a 3-D map to measure surface roughness. This work was done at the Indian Institute of Technology, Delhi with Dr. Kedar Khare.
Face recognition tool using deep learning
Face detection and face recognition using extremely few training examples were implemented using deep learning architectures. We used MTCNN network for cropping, aligning, and detecting faces. The face recognition was based on Face-Net network developed by Facebook and it used google's Image-Net architecture. This work was done at the Indian Institute of Technology, Delhi with Dr. Brejesh Lall.
Reconstruction from highly irregular samples
Reconstruction of an irregularly sampled, band-limited object with known bandwidth and less than Nyquist sampling. We exploited the theory of alternating projections onto convex sets (APOCS). We implemented the reconstruction by using both sinc and bessel basis functions. This work was done at the Indian Institute of Technology, Delhi.
