Optics-free Image Classification with Deep Metric Learning

Abstract

The lens is defined as a device to control how light reaches the imaging surface and is a fundamental component of imaging. Imaging applications are ubiquitous, ranging from autonomous driving to biomedical applications. With advancements in imaging technology, new applications in fields such as biomedicine and defense are driving a significant push toward the miniaturization of cameras. Unfortunately, this miniaturization has a fundamental difficulty: the total amount of light collected at the sensor image decreases with the lens aperture. As a result, ultra-miniature images collected simply by scaling down the optics and sensors are noisy. Thus, an innovative concept is introduced in which the lens is removed; however, the resulting images obtained without the lens are degraded. To address this issue, an alternative encoding scheme and computational algorithms are used to retrieve back the image, which we refer to as optics-free imaging. This thesis proposes a novel approach to using Deep Metric Learning for opticsfree image classification. Our Deep Convolutional Neural Network uses the image similarity metric for its learning algorithms. In this thesis, first, we trained our model with the dataset composed of degraded Cifar10 images taken in the lab and the original Cifar10 dataset for image classification. In the next task, we train our model on opticsfree (reconstructive) Cifar10 images obtained from degraded Cifar10 images using CycleGAN, along with original Cifar10 images for image classification. Finally, we employed Bayesian Prior to update the learning and computed the KL divergence.