Image Sanitization for Optimizing Video Dataset for Deep Learning | HCLTech

Deep Learning is expanding its footprint almost in all domains in creating intelligent solutions. The self-learning capability of the Deep Neural Network (DNN) enables perception-based solution which is very essential in many segments; example: autonomous vehicles, robotics and real-time video analytics predictive maintenance. Most of the AI enabled self-aware systems handles the image data from multiple image sensors. These image data could be of raw video data or loss-less or lossy compressed video data.

 To train the Neural Network for these applications itself requires a lot of real-time video data covering different scenarios. The dataset has to be balanced between the scenarios as well. The accuracy and reliability of the model majorly depends on the quality and distribution of dataset used for training and validation. The quality of image from different scenarios are essential to create a meaningful dataset database. In the process of data collection (i.e. video recording) the “image sanitizing” becomes important in removing the redundant image frames which gets recorded. Sanitizing the real-time video posts challenges in computing platform. This paper discusses storing of different image frames which creates major pain point in-terms of data storage and annotation.

Download the Whitepaper