Fig. 1. Pipeline of IEIM. (a) The IEIM data collection device. It employs a periodic pulsed modulation of light intensity with a period T. The collected events are processed according to the (b) inter-event interval (IEI) principle, where intervals between events reflect the intensity. (c) IEI calculations are performed on all event streams, and an image is selected from each cycle to represent the intensity at that specific moment.

Fig. 2. Comparison in synthesized static scenes. (a) Qualitative comparison results on normal synthesized data: membrane and mitochondria. (b) Qualitative comparison results on color-reversed synthesized data: actin and nucleus. Our method significantly outperforms SOTA methods.

Fig. 3. Comparison in synthesized motion scenes. Our method significantly outperforms SOTA methods. The red arrow in the GT image points to the area with significant motion. Scale bar, 5 μm.

Fig. 4. Comparison in real-world static scenes. (a) When the power is fixed, the imaging quality varies with the modulation frequency. Excessively high frequencies can lead to a decline in imaging quality, as evident from the zoomed-in details in the second row, indicated by the arrow. (b) When the modulation frequency is fixed, the imaging quality varies with the power. Increasing power can improve imaging quality, but too much power can cause detail loss, as evident from the zoomed-in details in the second row, indicated by the arrow. (c) The reconstruction results of SOTA models on data with a power of 6 mW and a modulation frequency of 500 Hz. The reference image is a region captured by the sCMOS camera that overlaps with the field of view of the event camera. Imaging with IEIM requires appropriate calibration between modulation frequency and light power, and the resulting image quality far exceeds that of SOTA methods. The power in the lower right corner corresponds to the laser power at the output of the objective lens. Scale bar, 2 μm in (a)–(c).

Fig. 5. Comparison in real-world motion scenes. (a) Dynamic imaging results of IEIM at a modulation frequency of 800 Hz and a light power of 9 mW. The second row of images shows detailed views of the white boxed areas in the first row of images, with the dashed lines indicating fixed positions. (b) Reconstruction results of E2VID at a temporal resolution of 1.25 ms. (c) Reconstruction results of E2VID at a temporal resolution of 0.62 ms. (d) Reconstruction results of E2VID at a temporal resolution of 0.31 ms. Our method achieves high-quality imaging with high temporal resolution without the trade-off between temporal resolution and reconstruction detail seen in methods like E2VID. The values in parentheses indicate the temporal resolution of the reconstruction, and the arrow points to the significant difference caused by the variation in reconstruction time resolution. (e) Intensity variation curve along the yellow solid line in (a)–(d). (f) Intensity variation over time at the white dashed line position in (a). Scale bar, 2 μm in (a)–(d).

Fig. 6. Dynamic imaging of common freshwater euglenae. The dynamic processes are presented at three different temporal resolutions, with a minimum of 2 ms, demonstrating that IEIM can effectively image dynamic processes. The images in the second row are magnified views of the regions enclosed by the white boxes in the first row, with dashed lines serving as reference markers to provide a more intuitive visualization of the motion.

Fig. 7. Comparison experiments with and without the modulation device. p.d.: pulsed modulation device. (a)–(c) Results in synthesized data. (d)–(f) Results in real-world data. The reference image is a region captured by the sCMOS camera that overlaps with the field of view of the event camera. Scale bar, 5 μm.

Fig. 8. Comparison of our method and existing network-based methods on synthesized static data with light modulation.

Fig. 9. Comparison of our method and existing network-based methods on real static data with light modulation. Scale bar, 2 μm.

Fig. 10. Comparison of the original network, the retrained network, and our method.

Fig. 11. Comparison of different methods on dynamic imaging of in vivo freshwater euglenae. (a) Reconstruction motion sequences spanning over 60 ms at temporal resolutions of 6 ms, 4 ms, and 2 ms. The arrow points to the fastest-moving part. (b) Intensity variation curve along the yellow solid line in (a). (c) Intensity variation over time at the red dashed line position in (a). Scale bar, 5 μm.

Fig. 12. Comparison of imaging performance between DAVIS346 and Prophesee EVK4 cameras. (a) Event accumulation map from the Prophesee camera over a 30 ms time window. (b) Event map and grayscale image from the DAVIS camera under the same conditions. (c) Event density maps generated by both cameras over a 1 s duration. The lower-left corners of (a) and (b) indicate the image resolution.