Grant: Discovery Grants Program - Individual

Signal processing theory and algorithms have a wide range of applications such as in image, video biomedical and speech processing. Traditional algorithms work best over structured small size databases, but are prone to erroneous results when data is scaled or when it is unstructured. On a small dataset, traditional algorithms do a great job, but once the data scale grows enormously, the large neural nets outperform others. This is partially due to the fact that a neural network model has many more parameters and has the ability to learn complicated nonlinear patterns. Deep learning algorithms have the ability to extract information at multiple levels in deep neural network architecture and map any data into a corresponding output class. Overall objectives of the proposed research are to develop a variety of efficient & reliable algorithms for processing image and video signals, using traditional as well as deep neural networks. In particular, it is proposed to concentrate over the next few years on video surveillance under hazy conditions and action-based video synopsis generation in a multi-camera network infrastructure. (a) A very critical issue in surveillance is perceptible scene understanding through detection and recognition for effective decision-making. Images captured in hazy environmental conditions suffer from poor visibility, low contrast and color cast issues. A unified framework that handles haze-induced distortions for effective detection and recognition of scene content is of utmost importance. Most of the existing approaches do not consider explicit mechanisms to handle the color cast during dehazing. It is important to consider whether a hazy image contains color cast before performing color correction. It is intended to build a cutting-edge AI technology, which upon deploying onto the system can automatically handle haze-related issues, aiming at effective scene analytics for surveillance and exploration. (b) Video surveillance is one of the most demanding areas of research in present day considering the unceasing security threats. Presently, more emphasis is being laid on video synopsis generation to make the surveillance automated and intelligent. The primary objective of synopsis is to produce a short summery of the video from a long one by simultaneous visualization of multiple objects in different time zones. In the past, video surveillance has been mostly limited to single static camera confined to indoor environment. Presently, multi-camera networks are used extensively, where the same scene is recorded from multiple angles. Mono-view video synopsis technologies cannot directly be applied to multi-view videos. Hence, this proposal aims at an action-based video synopsis generation using multi-camera network by establishing correspondence among multiple views and actions. (c) Finally, video synopsis would be combined with part (a) of the proposal to study the problem of synopsis of videos taken under hazy conditions.