The invention of digital holograms, advancements in computational technologies, and data-driven science led to research on digital holographic object recognition via machine learning methods. The methods proposed by this theses open a new paradigm for electromagnetic wavefront analysis, particularly at the optical frequency for digital hologram complex signals. Even though holography is an excellent solution for recording 3D images, an interference hologram consists of high-frequency fringe intensity images that are almost impossible to recognize through traditional computer vision methods. Traditional hologram recognition methods use optical correlation; however, this optical correlation can only result in a high correlation score if the two compared objects have similar poses and depth. All these call for the need for a method that can directly identify a digital hologram. This theses focuses on direct methods of analyzing object wavefront represented by a digital hologram complex signal’s magnitude and phase components through deep learning methods. Electromagnetic optical wavefront is in high complicacy due to its tangling information structure, as every single hologram pixel aggregate optical wave propagates from every point of the originating object. Conventional feature extraction technique by traditional 2D signal processing approach was not feasible to cover all variations’ combinations for invariant classification of deformable, occluded holographic objects with speckle noise interruptions. This study employs deep learning methods with automated end-to-end training and eliminates the whole human feature extraction process.

The new paradigm eliminates most of the above problems and is based on the widely accepted digital hologram complex signal format, which has been proved to be an effective representation of optical wavefront information. Nowadays, many research projects take off-the-shelf real number deep learning architecture to various holographic applications, and most are relatively heavy in model size. The proposed method is built from the ground up for complex signals and is based on the fundamental principle of linear regression representation through perceptron’s architecture. It works together with modern ReLU (Rectified Linear Unit) activation function, convolution filters, drop out regularization, Softmax decision unit, and various ways of building up hologram-classifier by ensemble magnitude component and phase component. Experiments prove that the method and structure effectively model complex signal information and provides an effective computation model for constructing decision boundaries to classify problems represented by complex wavefront signals. Consequently, unlike most optical correlation methods, the feature extraction process is fully automated during the deep learning network’s training process.

In addition to the latest version of the proposed method, EDL-IOHC Ensemble Deep Learning for Occluded Holographic Object Recognition, its two predecessors are also described in detail for ease of explanation and understanding. The EDL-IOHC can handle the problem of invariant classification on deformable objects with occlusion under noisy conditions in a single holistic framework. Furthermore, the EDL-IOHC is applied to multi-depths composite digit holograms. Followed by this, an extra application of it on a holographic interferometer is realized to further validate the theory and method for dealing with optically captured digital holograms instead of computer-generated holograms. The interferometer does not require an intact sample with a complete outline shape of the specimens or the organs to recognize the objects' identities. As the thickness of the specimens is several times longer than the wavelength of the laser employed, it is relatively a non-planar 3D object. The experiment proves that EDL-IOHC can also handle the situation well. The hologram-classifier EDL-IOHC can attain over 95% accuracy in recognizing occluded holographic handwritten digits objects under noisy conditions or biological specimens. It can be generalized from one domain to another domain almost effortlessly. Thus, we believe the proposing method EDL-IOHC can be taken as a reference framework and be extended to numerous disciplines such as industrial measurement, remote sensing and inspection, and medical imaging when complex signal wavefronts are involved. For example, electromagnetic wave spectrum frequencies at UV (Ultra Violet) and IR (Infra-Red) should be a direct application without significant modification, as off-the-shelf UV and IR image sensors at these frequencies are available in the market.