A vitrified protein particle has an unknown, effectively random orientation under the electron beam. Each shot integrates its density along the beam direction into one noisy 2D projection — real cryo-EM images sit far below unit SNR, so a single particle tells you almost nothing.
Once an orientation is assigned to each image (particle alignment/classification, assumed solved here), every projection can be smeared back along its own beam direction into a shared 3D voxel grid — back-projection. Averaging thousands of independent random-angle views cancels the noise while the true density, hit from every direction, reinforces itself.
- More particles → map correlation with the true structure rises, resolution number (Å) drops.
- Noisier images need proportionally more particles to reach the same resolution — the reason real datasets run into the hundreds of thousands of particles.