The absorption and scattering by water and suspended particles severely disrupt the optical field information during the propagation of light. Due to the extremely high demand for the integrity of optical field information in end-to-end imaging technology, image detail loss is easily caused during underwater imaging. Correlated imaging technology reconstructs the object image by calculating the correlation of optical field intensity fluctuations, significantly reducing the demand for detailed optical field information. However, traditional correlated imaging algorithms require the acquisition of a large amount of data from the object. Compressive sensing and deep learning algorithms can construct higher-order functions to nonlinearly analyze the optical field intensity fluctuations, thereby significantly reducing the number of data collection times for the object. Yet, the imaging results heavily rely on the quality of one-dimensional light intensity information collection. Therefore, how to efficiently utilize multidimensional information of the object in a strong absorption and scattering underwater environment remains an urgent problem to be solved.
Recently, Professor Li Xuelong's team from Northwestern Polytechnical University published a paper in Advanced Imaging titled "Multi-polarization fusion network for ghost imaging through dynamic scattering media", introducing an correlated imaging method based Multi-polarization Fusion mutual supervision Network (MPFNet). By optimizing information collection, information analysis, and network training, the method significantly improves the quality of reconstructed target images, providing an effective solution for deep learning-assisted underwater correlated imaging.
During the information collection phase, the method involves acquiring one-dimensional light intensity signals of the object either transmitted or reflected under the illumination of linearly polarized and circularly polarized lasers. This allows the retrieval of the object's multi-polarization information. In the information analysis phase, a multi-branch fusion architecture is designed to integrate and analyze the fluctuations of the signals collected under both polarization modes, thereby enhancing the efficiency of information analysis. In the network training phase, the multi-polarization information collected from object is used as the label, which constrains the convergence of network parameters and improves the generalizability of the method. This approach can effectively reconstruct high-quality object images in various strong scattering environments, including simulations, space, and underwater conditions.
Multi-polarization signal acquisition and fusion analysis: In the phase of information collection, one-dimensional object signals are acquired under the illumination of linearly polarized and circularly polarized lasers. Linearly polarized light can effectively filter out the backscattering from suspended particles in water, while circularly polarized light provides more stable signal transmission, reducing the degradation of light field information caused by scattering. The fusion of linearly polarized and circularly polarized light can further improve the completeness of the object echo signal.
Analysis and processing of object signal fluctuations: A multi-branch fusion architecture is designed to analyze linearly polarized and circularly polarized object signals at different scales. A multi-branch spatial channel cross-attention module is designed to explore the correlation between object linear polarization information and circular polarization information, promoting the fusion of multi-branch feature information between the encoder and decoder. The collaborative fusion of object linear and circular polarization signals significantly improves the fidelity of reconstructed object information.
Generalization of multi-polarization fusion processing: MPFNet fully incorporates the underlying physical model, using the acquired one-dimensional object signals as supervisory labels to avoid reliance on pre-training. Experimentally validated in free space and underwater environments, the MPFNet method significantly improves the correlation imaging effects in strong dynamic scattering media.
In summary, this paper presents a correlation imaging method based on multi-polarization information fusion supervision. By designing a multi-branch fusion network to integrate the wave characteristics of object linear polarization and circular polarization signals, and embedding the physical model into the MPFNet network, the obtained one-dimensional object signals are used as supervision labels to enhance generalization. Experimental results prove that by fusing the target signals of linear polarization and circular polarization, high-quality reconstruction of object images in different scattering media can be achieved. We believe that the proposed method provides an effective solution for deep learning-assisted correlation imaging in strong dynamic scattering media.
