A Hessian-Driven Convolutional Sparse Coding and Depth Awareness Algorithm for 3D Localization of Light Field Microscopy

Light-field microscopy (LFM), as an image acquisition device based on multi-sensor capturing of multi-view information, has gained widespread usage in the field of cell biology and neuroscience for the rapid and continuous recording of the spatial 3D distribution of cells and tissues. Presently, techniques utilizing deconvolution and deep learning are extensively employed for the analysis of data captured through LFM. Nevertheless, current deconvolution techniques continue to encounter constraints related to limited depth range, inadequate multi-source localization, and prolonged localization time, impeding their advancement. Moreover, deep learning methods are constrained in terms of their robustness and generalization ability. This paper introduces a novel approach to addressing the 3D localization challenge in LFM. This method is mainly based on a depth slice-based dictionary for more effective sparse coefficient solving and more accurate depth awareness. Specifically, a wave optics model is initially employed to create a slice-based dictionary comprising pre-constructed epipolar plane images (EPI) of the object at various depths. Subsequently, a convolutional sparse coding (CSC) approach is employed in conjunction with Hessian-driven damping to derive sparse coefficients for the EPI of the object within this dictionary. Then, a peak separation and clustering algorithm for energy aggregation is employed to separate and cluster multiple scenarios. Finally, an inversion verification algorithm is utilized to determine the optimal distribution of point sources. By evaluating our approach on simulated data and real-capture salina algae data, we have confirmed its efficacy in enhancing the depth range and localization accuracy, particularly for multi-point sources, while also decreasing the time required for localization in low signal-to-noise ratio (SNR) environments.