Computational holographic system for intelligent diagnosis

Abstract

In computational holographic imaging (CHI), the holographic interference patterns are recorded by a digital imaging sensor and processed by advanced computational methods. CHI performs significant advantages in these aspects, including high-resolution three-dimensional reconstruction, non-invasive and label-free imaging, real-time imaging, etc. The recorded holographic interference patterns enclose the structure and composition of the detected objects and can be used for their identification.

With advanced intelligent algorithms, such as deep learning and machine learning, the automatic analysis of holographic patterns can be achieved. However, there are several challenges currently existing to realize the intelligent diagnosis. First, the image data source is limited. The well-labeled dataset is difficult to acquire. Second, the interpretation between the physical model and networks is in the primary stage and needs to be further explored. Third, the amount of data and information generated by computational holography can be very large. Effective image processing methods are needed for rapid data analysis. This dissertation aims to alleviate the above-mentioned problems and provide competitive solutions with computational holographic imaging technology.

Firstly, we design a lightweight convolutional neural network (CNN) to deeply interpret and analyze the holographic data with a supervised learning-based method, termed a holographic classifier - convolutional neural network (HC-CNN). We experimentally show the effectiveness of the method in MPs classification and quantification. We also demonstrate its good performance in accuracy, robustness, and time efficiency.

Secondly, we creatively introduce a physical model (underwater image formation model) to specifically process the image descattering. The model-based method can get rid of the redundancy caused by the random initialization for the network training. We present the descattering results under different scenarios with single and complex image content. This method has good application significance in a wide range of forward and inverse imaging systems.

Thirdly, a transfer learning (TL) method as well as a class-based cross-entropy loss function and concatenated ReLU (CReLU) activation method is proposed for small image set classification. In addition, to achieve image augmentation effectively, a generative adversarial network (ACGAN) is developed. Multi-dimensional experimental results are demonstrated for the visualization of method effectiveness.

Fourthly, to relieve the challenges in acquiring the dataset and its labels, we design a zero-shot learning method to make use of the semantic attributes for image classification. By learning a mapping between the semantic features and the visual features of the objects, our method can recognize the objects that were not present during the network training process. This method is a powerful tool for image classification, especially when the labeled data is expensive to obtain.

Last but not least, we conclude with a summary of the main works in this thesis. We also perform a bright prospect for possible future working with advanced holographic imaging and machine intelligence.