Multiview sparsity quantification based class-manifold orthogonalization method for holography particle field characterization

In the study of high-speed particle fields analysis and characterization, the entire dynamic process is typically split into parts for description, with significant differences in particle density, size, velocity, and shape across different breakup regions. Digital holography, avoiding the cumbersome chemical processing and capturing instantaneous dynamics with high resolution and precision, has been widely used for high-speed particle field characterization. Currently, two primary approaches are normally employed for particle characterization: extended-focus-image and reconstructed 3D data slices. The former struggles with accurate segmentation under significant interference, while the latter meets challenges of low computational efficiency in particle fields with fewer particle existence. Moreover, practically existed class imbalance phenomenon caused by size differences and irregular shapes also severely impacts characterization accuracy. To address these issues, we propose the Multiview Sparsity Quantification based Class-manifold Orthogonalization (MSCO) method, featuring a two-step framework. In the first step, the Multi-view Sparsity Quantification Network (MSQNet) employs dimensionality reduction to extract particle-contained regions. The Grouped Feature Orthogonal Network (GFONet) in the second step locates the focal layers and morphologically characterizes particles using feature reorganization and grouped feature orthogonalization. The method is evaluated on four kinds of particle field data. Experimental results demonstrate that our proposed method outperforms existing algorithms in terms of computational time consumption, recall rate, segmentation accuracy, and generalization capability in high-density particle fields.