Proposed and developed multiple novel approaches for fusion of multi-band images (containing 200–250 different bands). Each of the proposed techniques is based on different framework- (a) signal
processing, (b) multi-objective optimization, and (c) probabilistic framework. This was a part of my PhD work.
More details.
High Dynamic Range Scene Generation
Proposed a novel an optimization-based technique for generation of an HDR image from multiple images taken by varying camera exposure. It does not require the knowledge of camera response function, or tone mapping operators.
Recipient of Honorable Mention Award at FUSION 2011.
Quality Assessment of Fusion Systems
Proposed new quality measures and several modifications to existing measures for efficient performance evaluation of image fusion techniques (especially for fusion of few hundreds of images). Proposed the notion of consistency of image fusion techniques.
Synthetic Depth-of-Field Effect
Worked on algorithms for producing a synthetic DoF effect for images captured on a handheld camera. Worked extensively on depth estimation, image segmentation, and blur effects.
Synthetic Depth-of-Field Effect for Video
Developed algorithms for a synthetic DoF effect for a video captured on a handheld camera. Proposed near real-time methods to preserve consistency across video frames.
Quality Assessment of Compressed Images
Proposed and developed algorithms to analyze the quality of compressed images. This is implicitly useful to evaluate different image compression schemes with respect to quality measures and applications.
Visual Search aka CBIR
Developed a content-based image matching/retrieval system. This was a hybrid system employing classical image processing and deep learning. The DL models were written and trained from scratch.
Computational Photography
Developed a completely new application to incorporate effects such as re-coloring, re-shadowing, re-texturing, etc. A user can select a target and source objects. This application provides a new look to target as per source.
QR-like Code
Developed an application to detect and decode a QR-like code with a photo captured using a low quality camera in a variety of lighting conditions. Also, it identified subtle features in the captured image.
Face Recognition
Developed a system to enroll and verify the user using face recognition. The system could handle reasonable variation in lighting and head pose + position of the user.
3D Scene Processing
Proposed and developed a faster image stitching algorithm— where images of the same scene were captured from different angles. The intended use was to provide 3D information for an AR framework.
Object Tracking
Developed a set of object tracking mechanisms. These differed in terms of their processing speed, complexity, ability to handle lighting variations, and ability to handle occlusions.
OCR
Developed a character recognition system for Roman/English text. The system's pre-processing could efficiently deal with varying lighting conditions, multiple shadows, motion blur, and noise. It included an optional use of application-specific super-resolution module.
Industrial QC
Built inspection systems for industrial quality check. These involved a hybrid mechanism (classical Computer Vision + Machine Learning) to detect, track, and analyze desired objects in industrial setup.
Scene Selection in Videos
Developed algorithms to identify specific scenes from a video (such as sport or movie). The objective was automated detection of video snippets; employed ML classifiers along with custom rules.
Industrial Machine Vision for Robots
Proposed a machine vision system to identify free space and best-possible path for a robot to move in a cluttered 3D region. Also, identified modifications in case the camera needs to be mounted on a robot.
Joint RGB+D Processing
Developed a system to provide a joint Color (RGB) and Depth information given 2 camera setup. The work included co-registration or calibration of two different cameras. It also included intelligent up-sampling of depth stream which was acquired at a lower spatial resolution.
Medical Image Enhancement
Developed a cascade of mechanisms to enhance images from medical modalities from Dicom. This contained noise suppression, contrast enhancement, and sharpening.
Artificial Sketching
Developed a novel algorithm to simulate a sketching animation. This algorithm analyzes the input image, and produces an animation video as if it was being sketched by hand. Additionally, it provides coloring where user can specify a palette of their choice. Some fun demos are here.
Stock Value Prediction
Developed a stock value prediction using Black-Scholes' formula. Modifications over the original formula included adaptive nature of weight parameters as well as wavelet decomposition for a better time-frequency analysis. This one is my favourite :-)
* The project list is selective, with a very brief description due to NDA or IP policies of corresponding companies or institutions.