Course Projects:

These are some old projects I worked on during my master's (2016-2018).

• Deformable Image Registration with a Convolution Neural Network

This project presents a Deep learning-based end-to-end network for deformable image registration (DIRNet). First, the network learns from moving and fixed images to output parameters for the spatial transformer, which generates the displacement vector field, and then warp the moving image to the fixed image using a resampler. It is a 2D non-iterative and unsupervised image registration method. DIRNet is evaluated on MNIST, Fundus, and Chest X-ray images.

• Template matching and Optical Marker Recognition (OMR)

This work involves aligning the marked OMR with template OMR. Then, recognise the marked entry and process the results. Used a Feature-based registration technique with similarity transform for aligning marked OMR. Then, by grid-wise matching with the template, entries are recorded to evaluate the results.

• An Ensemble Classifier for Rectifying Classification Error

An ensemble classifier for raw data using Assorter-Rectifier Method.  The assorter serves as the "base" classifier, which gives the initial prediction of the class. It is generally a parametric classifier. The rectifier would try to correct the prediction made by the assorter. It is generally a non-parametric classifier. As a part of this project implemented Tree Augmented Naive Bayes (TAN) algorithm for assorter and Logistic Regression for rectifier.

• Cascading Bandits: Learning to Rank in the Cascade Model

The Cascade Model is a popular model of user behaviour in web search. Model learns to rank the user clicks. Explored two algorithms in solving cascading bandits: Cascade UCB1 and Cascade KL-UCB. It is observed that CascadeKL-UCB performs better than CascadeUCB1 at the cost of runtime.

• NMF-KNN: Image Annotation using Weighted Multi-view Non-negative Matrix Factorization

The image annotation (assigning tags) has two significant drawbacks. First, models are relearned when new images and tags are added to the database. Second, the task of feature-fusion is typically dealt with ad-hoc approaches. To address this problem, presented a weighted extension of Multi-view Non-negative Matrix Factorization (NMF) based image annotation. The key idea is to learn a query-specific generative model on the features of nearest-neighbours and tags using the proposed NMF approach.

• Image Denoising using Discrete Shearlet Transform

Digital image captured by image sensors are generally affected by noise during the acquisition process and some time noise can be introduced during transmission and/or compression process. Denoising image with preserving edges are essential, especially in the medical domain. Hence, a Discrete Shearlet transform (DST) is used for denoising. It has multi-scale and multidimensional properties which are optimal in representing images containing edges.