Avinash

Avinash

If you help someone, you help everyone!

Exploring text summarization

less than 1 minute read

Published:

The problem we investigated was exploring and improving text summarization. Between generated and extracted summaries, our main focus was to improve the quality of generated (abstractive) summary. In particular we focused on building upon PEGASUS, a SOTA transformer model and investigate ways to improve the qualitative performance of abstractive summaries. Proposed and implemented Self-Attention Guided Copy Mechanism which guides the summarization model to copy the important source words from the source doc/article to the summary. You can find the project at this link

You May Also Enjoy

Discrete Wavelet Transform Image Super-Resolution Analysis and Deep Wavelet Prediction

less than 1 minute read

Published:

Super resolution techniques are useful in a range of industry applications from medical imaging and diagnosis to security and surveillance utilization. Since resolution enhancement of visual data from an imaging hardware point of view is costly, sensitive to environmental conditions, and time consuming, post-processing super resolution methods are a reasonable solution to improving the performance and analysis of imaging applications. In this paper, we analyze the performance of a modified Discrete Wavelet Transform (DWT) based super resolution algorithm on a variety of image acquisition types. A wavelet transform provides a detail as well as coarse separation of the contents of the image. We design a Convolutional Neural Network (CNN) to predict the missing details (residuals) in the wavelet coefficients of the low-resolution images to obtain high-resolution images. The network has multiple input and output channels which allows it to learn the different structures at different levels of the image. You can find more info here

Chest pathology classification in X-rays using GANs

less than 1 minute read

Published:

Medical datasets are often highly imbalanced with overrepresentation of common medical problems and a shortage of data from rare conditions. We propose simulation of pathology in images to overcome the above limitations. Using chest X-rays as a model medical image, we implement a generative adversarial network (GAN) to create artificial images based upon a modest sized labeled dataset. We employ a combination of real and artificial images to train a deep convolutional neural network (DCNN) to detect pathology across fiveclasses (Cardiomegaly, Pleural Effusion ,Pulmonary Edema, Pneumothorax and Normal) of chest X-rays.

Style transfer

less than 1 minute read

Published:

Humans have mastered the method of creating artistic images of different styles using methodology unique to them. Art and painting style is all about percep- tion and hence is considered a interplay of content and style of the scene to be described. Current world art has classified painting styles in different categories like Modernism, Figurative Art etc. Thus far the algorithmic basis of this process is unknown and there exists no artificial system with similar capabilities. However,advances in visual perception and recognition tasks using deep learning methods opens great opportunities for its extension to develop a system for style and contentlearning. This project attempts to understand the use of deep neural net architectures for learning and transfer of style and content of a given image. To know more vist link

Low cost Syringe pump

less than 1 minute read

Published:

Created a low cost syringe pump using a stepper motor and Arduino micro-controller. Fabricated the parts using a Mojo 3D printer. Enabled serial communication between Arduino and the computer using CoolTerm so that commands (for ex:- setting the flow rate) can be given by the user from terminal.This is a prototype of a syringe pump which can be used in applications where we require a fixed and precise fluid rate. A lot of credit goes to Prof. C. A. Varnon who has already created an open source syringe pump, from which we have adapted quite a few ideas. You can find the working demo of the prototype at this link