Scattering imaging of moving targets based on recurrent neural networks: a fusion of temporal and phase information

The speckle degradation caused by scattering media posed barriers to capturing the displacement information of moving objects. Moreover, the optical memory effect (OME) further restricts the imaging range, and current methodologies neglect the long-term temporal dynamics of the entire motion process. To overcome these limitations, we propose the temporal and phase recurrent neural network (TP-RNN), which leverages the temporal properties of recurrent neural networks (RNNs) and incorporates the phase extraction block (PEB) for handling phase spectrum information. The RNN unit captures and transmits temporal information between speckle image frames, while the PEB focuses on extracting key phase spectrum details from the features. This fusion strategy enables the network to recover moving objects and perform accurate quantitative predictions effectively. Experimental results validate that TP-RNN achieves high-precision imaging for moving targets across diverse trajectories and complexities, limiting prediction errors to within 1 pixel.