Fig. 1. Diagram of Hformer’s application scenario and its principle. (a) Schematic of the lensless holographic imaging system and the dynamic-static separation task; (b) structural diagram illustrating the principle of Hformer

Fig. 2. Hformer, which combines the LeWin module based on the Transformer architecture and adopts a dual-input and dual-output “H”-shaped structure, where the block number denotes the image’s channel count

Fig. 3. Dataset generation process

Fig. 4. The dynamic and static hologram segmentation results obtained from the stepwise ablation experiment, along with the corresponding 2D-shape reconstruction results. (a) Original hologram G_t to be segmented; (b) 2D-shape image directly reconstructed from G_t; (c) ground truth corresponding to Fig. 4(b); (d)‒(m) static and dynamic particle holograms segmented using different Hformer ablation variants and the complete Hformer, with the corresponding ground truth provided for comparison; (n)‒(w) 2D-shape images reconstructed from the segmented holograms

Fig. 5. Comparison of Hformer output results based on supervised learning and self-supervised learning

Fig. 6. Experimental system and results. (a) Schematic diagram of the experimental system; (b)‒(d) Original images G_t of pollen samples collected at three different time points t₁, t₂, and t₃; (e)‒(g) 2D-shape images reconstructed directly from the original images G_t at the three time points; (h)‒(j) dynamic particle holograms segmented using Hformer at the three time points; (k)‒(m) 2D-shape images reconstructed from the segmented dynamic particle holograms