Singular value decomposition in image processing

Abstract

The rapid advancement of object detection models in recent years has provided fast and accurate object detections on critical applications such as autonomous driving and healthcare services. While current-state-of-the art object detection models have achieved remarkable accuracy, it is at the expense of high computational cost and a significant amount of training time and data. In this project, we investigate various compressive sensing techniques, such as applying SVD compression and SVD background subtraction on the training images with the YOLOv5 object detection model in an attempt to investigate the benefit of compressive sensing on object detection tasks. The SVD-IC dataset, with SVD compression applied on the training dataset, outperformed the original dataset by achieving 7.9% higher mAP and 3.1% higher precision accuracy during testing. When SVD compression and SVD background subtraction were applied to the training dataset, it was observed that it enhances performance during the early stage of training. Systems with lower computational resources are able to benefit from SVD compression, where compressed training images would result in lower storage usage, as well as obtaining object detection models with better performance when trained with low number of epochs.