Computational Methods for Improving Fiber Bundle Imaging Systems

Abstract

Fiber bundle imaging systems use a bundle of single coherent fiber cores as imaging probe to capture object information from remote areas. They have the advantage of great flexibility and compact size. However, such imaging systems typically have two major limitations, namely honeycomb-like fixed patterns and limited spatial resolution. To address these two issues, we propose three new methods that will be suitable under different application scenarios. Our first method can jointly improve spatial resolution and remove fixed structural patterns for coherent fiber bundle imaging systems, based on inverting a principled forward model. The forward model maps a high-resolution representation to multiple images modeling random probe motions. We then apply a point spread function to simulate low resolution figure bundle image capture. Our forward model also uses a smoothing prior. We compute a maximum-a-posteriori (MAP) estimate of the high-resolution image from one or more low-resolution images using conjugate gradient descent. Unique aspects of our approach include (1) supporting a variety of possible applicable transformations; and (2) applying principled forward modeling and MAP estimation to this domain. We test our method on data synthesized from the USAF target, data captured from a transmissive USAF target, and data from lens tissue. In the case of the USAF target and 16 low-resolution captures, spatial resolution is enhanced by a factor of 2.8. Our second method is a deep learning based image restoration method, which can remove honeycomb patterns and improve resolution for fiber bundle images. By building and calibrating a dual-sensor imaging system, we capture fiber bundle images and corresponding ground truth data to train the network. Images without honeycomb patterns are restored from raw fiber bundle images as direct inputs, and spatial resolution is significantly enhanced for the trained sample type. We also construct the brightness mapping between the two image types for effective use of all data, providing the ability to output images of the expected brightness. We evaluate our framework with data obtained from lens tissues and human histological specimens using both objective and subjective measures. Our third method estimates high-resolution images from multiple fiber bundle images using deep learning. Our two-stage approach first aligns input frames with motion estimation neural network and then uses a 3D convolution neural network to learn a mapping from aligned fiber bundle image sequences to their corresponding ground truth high-resolution images. Evaluations on lens tissue data and USAF resolution target data demonstrate that our proposed deep learning framework can significantly improve spatial resolution.