Tools:- [Cudnn, Cuda, Python, Anaconda, Pytorch, Tensorflow, Kears, OpenCV, Numpy, Pandas]
Develop a novel lightweight deep learning (VS-Net) model for detecting and segmenting the document in the videos.
The VS-Net is designed with the help of a variational autoencoder with a depth-wise separable convolution network and approximation rank pooling.
The VS-Net extracts the multi-scale spatial and temporal information locally and fuses globally while preserving the temporal information using approximation rank pooling.
Evaluated and compared with existing state-of-the-arts methods in terms of network complexity and speed as well as peformance in accuracy.
The VS-Net is used in IDRBT for the Video KYC project.
Tools:- [Cudnn, Cuda, Python, Anaconda, Pytorch, OpenCV, Numpy, Pandas]
Develop a novel lightweight deep learning (DSNet) model for detecting and segmenting the object in the wild.
Designing new modules such as the Deformable Convolution Network (DeCNet), Separable Convolution Network (SCNet), and Depth-wise Attention Response Propagation (DARP) module.
The DSNet extracts the geometric multi-scale spatial and temporal features and fuses them to detect the objects efficiently while preserving the temporal information using DARP.
It evaluated and compared with existing state-of-the-art methods in terms of network complexity and speed as well as performance in accuracy.
The DSNet is sponsored by IDRBT for the Video KYC project.
Tools:- [Cudnn, Cuda, Python, Anaconda, Pytorch, OpenCV, Numpy, Pandas]
Developed a novel lightweight deep learning (HSNet) framework for detecting and segmenting shadows in the wild using geometric multi-scale features.
The HSNet extracts the multi-scale geometric low-level followed by high-level spatial and temporal information in a hierarchical way using various proposed modules (MFEM, GAIM, SGFM, SGM, etc.).
In the HSNet, the edge low-level information is extracted by the proposed Edge Detection Network (EDNet) and guides the HSNet for detecting and segmenting the geometric variation of shadows in the wild.
It is evaluated and compared with existing state-of-the-art methods in terms of network complexity and speed as well as performance in accuracy.
It is sponsored by IDRBT for the Video KYC project.
Tools:- [Cudnn, Cuda, Python, Anaconda, Sklearn, Tensorflow, Kears, OpenCV, Numpy, Pandas]
Develop an unsupervised deep learning (VS-Net) model for detecting whether card transactions are fraudulent or not.
In this, an autoencoder is used to encode the features in an unsupervised way, and a support vector machine (SVM) classifier is used to classify whether the transactions are fraudulent or not.
A oversampling and SOMTE class imbalance techniques are used to balance the class and compared with existing machine learning methods such as decision tree, random forests, and logistic regression.
It is evaluated on confusion matrix, accuracy, F1 score, and precision and recall.
Tools:- [Cudnn, Cuda, Python, Anaconda, Sklearn, Tensorflow, Kears, OpenCV, Numpy, Pandas]
Develop an unsupervised framework to predict anti-money laundering using deep convolution network (DCN) followed by support vector machine (SVM).
In this, DCN extracts features in an unsupervised way and pass to a support vector machine (SVM) to predict anti-money laundering.
It compared with existing machine learning methods such as decision tree, random forests, and logistic regression and autoencoder, which shows outstanding performance.
The performance is evaluated on using confusion matrix, accuracy, F1 score, and precision and recall.
The undersampling and oversampling techniques are used to balance the class as a preprocessing.
Handled various theoretical and practical session on e-programmes and offline programmes such as "Anti-money Laundering in Banking, Fraud-detection , and Expalinable AI in Banking orgainzed by IDRBT.
Handled hands-on session on various offline programes held during my research tenure such as "Real-time object detection and tracking in videos", "Fraud-detection using machine leanring models in banking", "Anti-money laundering prediction" held by IDRBT
Handled hands-on session and theories class for machine learning & aritificial intelligence and computer vision courses to PGDBT at IDRBT.