Research
All my research manuscripts are freely available through my Google Scholar or ResearchGate pages.
ReViewNet: A Fast and Resource Optimized Network for Enabling Safe Autonomous Driving in Hazy Weather Conditions
ReViewNet is a fast, lightweight and robust dehazing system suitable for autonomous vehicles. The network uses components like spatial feature pooling, quadruple color-cue, multi-look architecture and multi-weighted loss to effectively dehaze images captured by cameras of autonomous vehicles. The network's superiority was benchmarked on 5 datasets with a quantitative, qualitative and ablation analysis. (IEEE Transactions on Intelligent Transportation Systems July 2021).
CVPR NTIRE 2020 challenge on non-homogenous dehazing
This paper reviews the NTIRE 2020 Challenge on Non-Homogeneous Dehazing of images (restoration of rich details in hazy image). The report shows solutions of the top 25 teams in the world who made it to the leaderboard of the submission. It was wonderful to collaborate with some of the best computer vision researchers in the world and be a part of this conference's proceedings. (Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops 2020)
Smart water conservation through machine learning and blockchain-enabled decentralized edge computing
In this paper, we propose a blockchain based incentivized edge computing framework for water saving using soft computing methodologies. By using Feed Forward Networks and Mixture Density Networks, we predict the water usage in terms of input factors and historical usage respectively, thus incorporating machine computing into the framework. The blockchain and soft computing hybrid model ensures that accurate incentives are provided to the people in order to motivate them to avoid wastage of water. (Elsevier Applied Soft Computing 2021)
This is the world's first deep learning-based semantic segmentation model for land cover classification of temporal microwave images. Applying six deep learning architectures viz. Pyramid Scene Parsing, UNET, DeepLabv3+, Path Aggregation Network, and Feature Pyramid Network over temporally acquired SAR datasets for three different frequencies. Outputs of all six architectures have been assessed using frequency weighted IoU (Geocarto International Jan 2022).
TheiaNet: Towards fast and inexpensive CNN design choices for image dehazing
This work examines inexpensive design choices for dehazing as an end-to-end image-to-image mapping problem enabling dehazing in a highly resource constrained environments. The proposed TheiaNet beats existing methods in term of accuracy, speed, compute and memory efficiency highlighting how judicious application-specific components can augment simple CNNs to denoise faster, and more accurately than heavier networks, which is supported by an ablation analysis. (Elsevier Journal of Visual Communication and Image Representation 2021 )
A machine learning and blockchain based secure and cost-effective framework for minor medical consultations
In this paper, we present an advanced protected blockchain-based consultation model for minor medical conditions. This is done by using word2vec and TF-IDF weighting to classify the question and cosine similarity scores for detailed orientation analysis. Based on this score, the patient is charged, and simultaneously, the responder is awarded ether. An incentivized method leads to more accessible healthcare while reducing the cost itself. (Elsevier Sustainable Computing 2022)
DeRainGAN: Single image deraining using wasserstein GAN
In this paper, we design a simple yet effective ‘DerainGAN’ framework to achieve improved deraining performance over the existing state-of-the-art methods. The learning is based on a Wasserstein GAN and perceptual loss incorporated into the architecture. (Multimedia Tools and Applications 2021)
A review of text classification - fastText and DPCNN
This review paper covers 2 papers on the problem statement of general text classification. The first-fastText-is based on the method proposed by Facebook, and is still considered one of the fastest and most scalable methods for the task, without compromising on accuracy. The second-DPCNN-is proposed much later, and works on the word level to extract knowledge through the CNN flow. It proposes pre-activation and constant feature downsampling to give rise to a pyramid-shaped computation efficient model.