A novel approach that combines the advantages of neural networks with handcrafted features to perform face liveness detection against 2D spoofing attacks. We created a novel Flash Detection dataset for the task that includes images captured under variable flash and distance settings with five types of spoofing attacks.
Implemented a Siamese Bidirectional LSTM network with a combination of handcrafted features. Obtained benchmark results for the task on the Mohler dataset.
Designed and implemented a chatbot application for the COEP website to respond to queries about faculty, college fests, and campus buildings. Used tools such as Stardog for data warehousing, and LSTM cells for model training.
Generated Hindi word embeddings using a Negative Sampling Architecture. Utilised the embeddings along with a tf-idf scoring scheme to create a document ranking system.
