Abstract
Traditional deep-learning techniques for image reconstruction often demand extensive training datasets, which might not always be readily available. In response to this challenge, methods not requiring pre-trained models have been developed, leveraging the training of networks to reverse-engineer the physical principles behind image creation. In this context, we introduce an innovative approach with our untrained Res-U2Net model for phase retrieval. This model allows us to extract phase information, crucial for detecting alterations on an object's surface. We can use this information to create a mesh model to represent the object's three-dimensional structure visually. Our study evaluates the effectiveness of the Res-U2Net model in phase retrieval tasks, comparing its performance with that of the UNet and U2Net models, specifically using images from the GDXRAY dataset.
Fig.1. 3D phase retrieval: (a) 2D Ray-X test image, (b) 2D phase retrieval estimate, and (c) resulting 3D mesh.
Method
Overview of the proposed architecture for Phase Retrieval:
Fig.2. Res-U2Net architecture: (a) U2Net model configuration, based on a multi-scale sequence of Res-UNet models, (b) Res-UNet model, the encoder extracts features using convolutional layers (Conv2D) with batch normalization, ReLU activation (ResBlock), and spatial resolution reduction via max pooling (MaxPooling2D). This is followed by a decoder assigning phases to the features by upsampling using transpose convolutions (Conv2DTranspose) with skip connections. Residual connections link the encoder and decoder layers to improve the training performance. Finally, a 1×440×4401×440×440 convolutional layer generates the segmentation mask, resulting in the network output.
BibTeX
@article{OsorioQuero:24,author = {Carlos Osorio Quero and Daniel Leykam and Irving Rondon Ojeda},journal = {J. Opt. Soc. Am. A},keywords = {Biomedical imaging; Computational imaging; Fluorescence lifetime imaging; Imaging techniques; Inverse design; Phase retrieval},number = {5},pages = {766--773},publisher = {Optica Publishing Group},title = {Res-U2Net: untrained deep learning for phase retrieval and image reconstruction},volume = {41},month = {May},year = {2024},url {https://opg.optica.org/josaa/abstract.cfm?URI=josaa-41-5-766}, doi = {10.1364/JOSAA.511074}}
Comments