Ashwin Umadi

Ashwin Umadi

Master of Science in Computer Science | Ex - Senior Software Engineer at Sandvine Technologies

Publications

Pictorial Information Retrieval from EEG using Generative Adversarial Networks

Published in 2022 7th International Conference on Intelligent Informatics and Biomedical Science (ICIIBMS), 2022

Presented a supervised methodology to reconstruct images of various shapes and colors as seen by a person from brain signals. Generative AI and Computer Vision research, using an Emotive Enobio headset to obtain the EEG signals from the brain, and extracted the features using a CNN model. Finally, reconstructed these signals into images using conditional Deep Convolutional Generative Adversarial Networks(cDCGAN) models with an inception score of 12.67 (accuracy close to State of the Art models). Received the Best paper award in the International Conference on Intelligent Informatics and Biomedical Sciences, Japan 2022.

Recommended citation: K. Nagarajan, A. Umadi, N. Belur Keshav and N. Krupa, "Pictorial Information Retrieval from EEG using Generative Adversarial Networks," 2022 7th International Conference on Intelligent Informatics and Biomedical Science (ICIIBMS), Nara, Japan, 2022, pp. 269-275, doi: 10.1109/ICIIBMS55689.2022.9971471. keywords: {Visualization;Image color analysis;Convolution;Predictive models;Brain modeling;Feature extraction;Generative adversarial networks;EEG;Image reconstruction;Deep Learning;Brain Media;Generative adversarial networks(GANs)}, https://ieeexplore.ieee.org/document/9971471

Automated Segmentation of Overlapping Cells in Cervical Cytology Images Using Deep Learning

Published in 2020 IEEE 17th India Council International Conference (INDICON), 2020

Led the team of three and proposed a novel three stage approach for cell mass, nuclei and cytoplasm segmentation for overlapping cervical cells via Computer Vision techniques, involving a supervised Machine learning U-net Convolutional Neural Network. Training criterions used were Binary cross entropy loss function, Adam optimizer. First stage was Cell Mask segmentation, second Nuclei segmentation, and third stage included sending the previously segmented images as inputs along with each Cell’s cytoplasm into the U-Net model to obtain the contours of all overlapping cells. Achieved mean IoU metric of 90%, overperforming the contemporary state of the art techniques. The system was deployed on Google Colab GPUs to speed up the learning process.

Recommended citation: A. Umadi, K. Nagarajan, J. B. Venkatesha, A. Ganesh and K. George, "Automated Segmentation of Overlapping Cells in Cervical Cytology Images Using Deep Learning," 2020 IEEE 17th India Council International Conference (INDICON), New Delhi, India, 2020, pp. 1-7, doi: 10.1109/INDICON49873.2020.9342328. keywords: {Deep learning;Image segmentation;Visualization;Technological innovation;Shape;Conferences;Hardware;Overlapping cell segmentation;cervix cytology;K-means;U-Net}, https://ieeexplore.ieee.org/document/9342328