Programmable Spatially Variant Single Pixel Imaging Based On Compressive Sensing

Shin, Zhen Yong (2021) Programmable Spatially Variant Single Pixel Imaging Based On Compressive Sensing. Master dissertation/thesis, UTAR.

Preview

PDF Download (4Mb) | Preview

Abstract

Single-pixel imaging techniques made imaging easier in conditions that are unfavourable to the conventional digital cameras such as the invisible wavelengths. According to Nyquist-Shannon theorem, it requires that the number of measurements must be no less than the number of image pixels for an error-free image recovery. However, acquiring more measurements in practice increases the cost and operating time which hinder the practicality of single-pixel imaging. Single-pixel imaging system based on compressive sensing (CS) makes it possible to simultaneously compress and acquire image data, thus recovers images from measurements less than the requirement stated by Nyquist-Shannon theorem. In general, the image quality is proportional to the number of measurements which contributes to the acquisition and computational time accordingly. Hence, the essential goal of efficient singlepixel imaging is to maintain a high recovered image quality while reducing the number of measurements and processing time. In the conventional uniform resolution (UR) single-pixel imaging, all image pixels have the same size and are equally weighted. Therefore, a high pixel-to-pixel fidelity image recovery requires many measurements. However, high pixel-to-pixel fidelity recovery is not always needed in most visual tasks. This thesis proposes a programmable spatially variant resolution (SVR) technique in single-pixel imaging based on CS. In the proposed method, image iv pixels are differently sized and formed higher and lower resolution regions. Since lower resolution regions require less measurements for recovery, most of the measurements are spent in the recovery of the higher resolution regions. Thus, SVR single-pixel imaging is able to maintain higher image quality with significantly fewer measurements. Since recovering large images requires longer time and more computational resources, block-based CS technique is proposed to reduce the computational cost by dividing them into small image blocks. Each image block is recovered in the same way as recovering individual images. In this project, a design of SVR sensing patterns is proposed and applied to the block-based CS which can reduce the complexity and time needed for the result computation. Recently, many convolutional neural networks (CNN) were proposed as the CS image recovery algorithms. Conventionally, the network inputs of an CNN are images and the network outputs are the predictions of the labels of the images. In the framework where CNNs are applied as the CS image recovery algorithms, the network inputs would be the CS measurements and the network outputs would be the recovered images. The iterative nature of the conventional CS image recovery algorithms increases the computational resources needed and the time for image recovery. In comparison, CNNs are non-iterative. Many studies had shown that CNNs improve the recovered image quality and reduce the time cost for image recovery drastically. In this project, a CNN called ReconNet is adapted as the CS image recovery algorithm. The results of this project show that the proposed SVR sensing patterns are able to improve the image quality and time efficiency for small number of measurements as compared to the conventional UR sensing patterns. The improvements can be seen in both the conventional CS approach as well as the blockbased CS approach. the image quality In addition, SVR sensing patterns are able to retain better as the number of measurements gets smaller. Furthermore, the results have also shown that compared to the other conventional CS image recovery algorithms, ReconNet significantly reduces the time needed for image recovery small numbers of measurements Hence , the proposed . while maintaining a high image quality for SVR approach highfidelity pixeltois more suitable in situation s where pixel recovery is not the priority and most importantly far fewer measurements are required for a comparable image quality. ReconNet Moreover , outperforms the conventional CS image recovery algorithms. This shows that the pro posed SVR approach with ReconNet is more suitable than the conventional approaches for practical cases.

Actions (login required)