Research
Image borrowing for Improved Face Recognition
Lack of face images with wide variety in illumination is a huge constraint which limits the accuracy of any face recognition algorithm. Here we have proposed to borrow the illumination from one face to improve the same for the other. In the adjacent image, the face of Person1 is illuminated using the lighting conditions borrowed from Person2. A publication based on this work is submitted for review in ICIP 2015.
Content Detection in Cheque Images
Input Image
(Personal details blackened out)
Processed Image
This project involves automatic registration of cheque images and the detection of the various filled fields in it. One of the main constraints of this project is to implement it without using any of the currently patented algorithms (e.g. SIFT, SURF etc.). A template image in the form of a blank cheque of same bank is given in order to get the relative locations of the fields in the cheque. The output image is the registered cheque image with empty fields marked in RED and the filled ones in GREEN.
Object Tracking for Surveillance Applications
Indoor - Single Object
Outdoor - Distant Object
Indoor - Multiple Object
Outdoor - Fast Moving Object
The aim of this project is to track an object across multiple non-overlapping cameras in real-time. Each camera is connected to an embedded module, which consists of a Beagleboard-xM connected to the network via Ethernet. All the computations required for tracking are done on the embedded module. The data (containing the tracking information) is then sent to a central display system. A modified CAMShift algorithm, suitable for the system, is used for tracking. Apart from that, the implementation in C is also optimized in order to get better frame rate.
The system is tested for various conditions like occlusion, presence of multiple objects etc. The performance remains unchanged across indoor and outdoor scenarios. As seen from the images, it is able to track multiple objects, as well as fast moving objects in real-time. A frame rate of over 10 fps is achieved on Beagleboard-xM while tracking. The frame rate shoots up to 53 fps when implemented on a standard PC. In all the cases the resolution is set to 640x480 pixels.
This work is a part of my Masters thesis and is published in ISM 2013 at Anaheim, USA.
Automized Measurement of Frontal Plane Knee Alignment
Input Image
Processed Image
The aim of this project is to measure the Hip-Knee-Ankle Angle of the human leg from X-ray images. An edge image is obtained using Canny edge detector. Hough transform is then used to detect the straight lines in this image. The implementation is done using MATLAB. The angle is measured with a mean-square-error (MSE) of 0.35 degrees. In the image shown, the thick WHITE lines are detected using the algorithm. The RED and BLUE lines are drawn manually, only for the purpose of illustration.
This project is a part of my final year thesis during Bachelors degree.