Image Processing Scheme For Single Pixel Imaging

Ng, Jian Hau (2021) Image Processing Scheme For Single Pixel Imaging. Final Year Project, UTAR.

Preview

PDF Download (2051Kb) | Preview

Abstract

Single pixel imaging is substantially different from conventional imaging approaches which it needs only a bucket detector to capture images instead of a pixelated detector. It has been a promising method for imaging with nonvisible light, imaging through turbid media and weak-light conditions. Compressed sensing is commonly used in single pixel imaging to efficiently acquire and reconstruct signals with less than Nyquist sampling rate. Same resolution grade of masks formed by Pseudorandom or Hadamard matrix are used to sample information from the target scene. Image can be reconstructed by computational means based on the sample measurements. However, this technique is not adequate to produce a high-quality image which remains the major drawback. Therefore, this project aims to propose a suitable image processing scheme to achieve better image quality. Two approaches are proposed to improve the overall image quality; a Coarse-to-Fine sampling method in the data acquisition phase and Super Resolution enhancement in the post-image reconstruction phase. To realize the Coarse-to-Fine sampling, a sequence of masks with increasing resolution grade from low to high are used to capture different outline information. As for Super Resolution enhancement, a network formed using deep learning approach named Very Deep Super Resolution (VDSR) is used to estimate the information loss in the low resolution image. In general, the proposed Coarse-to-Fine sampling and VDSR method show improvement in image quality. Integrated scheme by combining both proposed methods is proven to outperform the conventional compressive sensing especially from the perspective of structural similarity index measurement (SSIM).

Actions (login required)