Computational imaging and detection via deep learning

Abstract

Data-driven signal and data modeling has received much attention recently, for its promising performance in image processing, computer vision, imaging, etc. Among many machine learning techniques, the popular deep learning has demonstrated promising performance in image-related applications. However, it is still unclear whether it can be applied to benefit various computational imaging and vision applications, ranging from image reconstruction to analysis. This dissertation gives a comprehensive overview of the fundamentals of deep learning for object detection, including logistic regression, forward propagation, backward propagation, optimization techniques (e.g., dropout, momentum, and Adam), convolutional neural networks and computer vision applications, with a glace at some advanced topics (e.g., bounding box prediction, non-max suppression, and region proposal). Some popular deep learning models, such as the LeNet-5, AlexNet, VGG-16, ResNet, and Inception, are discussed in detail. Focusing on the object detection task, this dissertation investigates the ideas and procedures of the YOLO algorithm in particular and presents implement details of a detection problem with X-ray images. Specifically, the X-ray images are fed into deep neural networks to predict the classes and locations of five types of dangerous items. We present experimental results showing the effectiveness of the implemented algorithm for detecting objects from X-ray images, towards building a fully automated security inspection system using deep learning and computer vision techniques.