Model compression of deep neural network architectures for visual pattern recognition: Current status and future directions

Abstract

Visual Pattern Recognition Networks (VPRNs) are widely used in various visual data based applications such as computer vision and edge AI. VPRNs help to enhance a machine's learning by extracting underlying patterns from input images and videos. VPRNs can be implemented using statistical, syntactical, or Machine Learning (ML) based approaches. With the popularity of Deep Neural Networks (DNNs) in the recent past, DNN based VPRNs have become an inevitable choice due to their suitability for handling such high dimensional data.However, such DNN based VPRNs bring along a curse of dimensionality, leading to intricate computations, substantial memory demands, and increased energy requirements. This impedes their practical deployment in resource constrained and strict latency required environments. Such overheads impose a demand for compression of VPRNs without impairing their performance. This research presents an exhaustive survey on compression techniques used for sequential, non-sequential and advanced DNN architectures from the perspective of Visual Pattern Recognition (VPR). Research findings of this study reveal potential scope to compress non-sequential, sequential and advanced DNN architectures for VPR applications. The study also clarifies challenges, reported and unreported issues as well as research gaps and provides directions for further research in this domain.