Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Electronic and Information Engineering | en_US |
| dc.contributor.advisor | Chen, Wen (EIE) | en_US |
| dc.contributor.advisor | Yu, Changyuan (EIE) | en_US |
| dc.creator | Zhou, Lina | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/11394 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Learning based optical imaging through scattering media and its applications | en_US |
| dcterms.abstract | With tremendous progress in optical imaging systems, researchers attempt to better implement imaging through scattering media. Imaging through scattering media is a long-standing problem which has raised great concerns in many fields, such as biology, biomedical science and marine ecology. The common difficulties encountered in imaging through scattering media are usually complexity and uncertainty of the environmental media. Existing methods developed for reconstruction of original images from speckle patterns are applied in a stable and pure medium, which obstruct the wider applications in unpredictable and complex scattering media. Moreover, extant solutions suffer from huge computational load, low resolution and limited application ranges, which cannot be effectively applied to resolve homologous imaging problems in scattering media. Advanced optical imaging technologies play an important role in further exploration of optical imaging in complicated environments, especially in strongly and randomly scattered media. Hence, it is essential and meaningful to take complex environments into consideration for better fulfillment of optical imaging. Fortunately, recent advances in machine learning inject fresh blood into optics leading to revolutionary improvements, which largely promote developments of optical imaging through complex and uncertain scattering media. This Ph.D. thesis begins with the study of optical imaging in scattering media (e.g., turbid water). Here, a learning based approach for optical imaging is proposed to effectively reconstruct high-quality objects in the turbid water for the first time. By exploiting certain image pairs of intensity images captured by a CCD camera and the corresponding input images sent to complex environment, a convolutional neural network (CNN) for imaging system can be well trained, and then ensures the distorted images to be reconstructed in real time. Moreover, it has a precedent advantage compared to conventional methods. The proposed method is robust to multi-range of turbidity which validates feasibility and great potentials for high-performance sensors in harsh environments, allowing unknown perturbations. In consideration of the changeable environments, a new approach based on cosine similarity for speckle classification and CNN for image reconstruction is proposed to effectively restore the object. The combined model is tolerant to the uncertainty of turbidity and also guarantees high-accuracy pattern classification and high-quality image reconstruction. In addition to the ability to predict objects in turbid media with uncertain densities, it is also verified that the proposed CNN model can be applied to recover objects placed in scattering media with arbitrary distances and arbitrary densities. In practice, accurate position of the object and accurate density of the turbid media are both uncertain which may greatly decrease the applicability of previous imaging techniques. In this study, the variants (e.g., density and distance) which seriously affect the reconstruction quality in water solutions have been systematically investigated. Although the accurate location and density are not known, it is still feasible to reconstruct the speckle patterns based on the proposed learning method. | en_US |
| dcterms.abstract | In view of feasibility and efficiency of the learning based method for image reconstruction through scattering media, applications of learning based attacks for optical cryptosystems are explored for the first time. Compared with conventional optical cryptoanalysis methods, the proposed learning based attacks can effectively retrieve original plaintexts without extracting optical security keys and using complex phase retrieval algorithms. This learning based attacking method is verified to be applicable to study the vulnerability of different optical encryption systems based on diffraction, interference and computer-generated holograms. Moreover, it is experimentally verified that the proposed method is also applicable for optical cryptosystems based on complex scattering environments with cascaded masks. It is expected that the proposed learning based attacking method can provide a new and effective means for optical cryptoanalysis, which will urge more secured encryption systems to be explored. Inspired by striking characteristics of machine learning and its excellent applications in optical cryptoanalysis, learning based optical encryption in complex scattering media is proposed for the first time to enrich optical cryptosystems. Traditionally, only when correct keys are used in the decryption process, original plaintext can be successfully retrieved. However, the learning based attack poses a great threat to the security of optical cryptosystems in which the plaintext can be directly retrieved without usage of optical security keys. Hence, it would be desirable to develop new techniques for optical cryptography. Instead of employing optical parameters used in the encryption as security keys, the proposed method makes use of training parameters used in machine learning models to serve as security keys. Except for the parameters used to train machine learning models, virtual phase-only masks are also applied to enlarge security space and improve information security. A series of sensitivity analyses have been conducted to verify effectiveness of the proposed method. It is experimentally verified that the proposed learning based optical encryption possesses high security, which can escape the invading of extant attacks. Furthermore, object authentication is studied and applied to further enhance system complexity of learning based optical encryption. The plaintext is processed by an extra security layer to be authenticated. Receivers need to make use of optical analytical tools to determine authenticity of the retrieved image obtained by learning based optical encryption. It is experimentally demonstrated that learning based optical encryption with object authentication dramatically achieves the higher security, which breaks new ground for optical security fields. It is expected that the findings could contribute to the evolutionary development of optical encryption schemes. In summary, this Ph.D. thesis presents the studies on learning based image reconstruction through scattering media, providing effective methods to solve complicated problems in optical imaging and expounding their applications in relevant research fields. It is hoped that this research work can help better understand applications of machine learning in optical imaging, and more research ideas can be stimulated to apply machine learning to optics and photonics from different perspectives. | en_US |
| dcterms.extent | xxvii, 179 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2021 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Imaging systems | en_US |
| dcterms.LCSH | Image processing -- Digital techniques | en_US |
| dcterms.LCSH | Optical data processing | en_US |
| dcterms.LCSH | Machine learning | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/11394