Physical Film Defect Detection and Restoration (Dust, Scratches, Surface Debris)

Topaz Video Ideas
1

Proposal for a New AI Model in Topaz Video AI: Physical Film Defect Detection and Restoration (Dust, Scratches, Surface Debris)

1. Executive Summary

Film‑based content (8mm, 16mm, 35mm, archival prints, telecine transfers) often contains physical surface defects such as dust, scratches, fibers, abrasion marks, and emulsion damage. These defects are not part of the photographed scene, but rather imperfections located on the surface of the film strip itself.

This distinction is crucial:

  • Physical defects cast micro‑shadows and exhibit a subtle 3D relief effect.

  • Real objects captured in the scene (birds, debris, distant shapes) do not cast shadows onto the film surface.

Current Topaz Video AI models (Proteus, Artemis, Iris) are not designed to detect or remove these film‑surface artifacts, as they focus on noise, grain, compression defects, and detail enhancement.

A dedicated model — or a pre‑processing inference step — could accurately detect and remove these physical defects by analyzing their optical relief, temporal persistence, and non‑photographic characteristics.

A dedicated model — or a pre‑processing inference step — could accurately detect and remove these physical defects by analyzing their optical relief, temporal persistence, and non‑photographic characteristics.

2. Nature of the Problem

2.1. Types of Physical Film Defects

  • Dust particles trapped between film and gate

  • Vertical or horizontal scratches

  • Emulsion abrasion

  • Fibers or hair stuck on the film

  • Chemical stains or residue

  • Pressure marks from mechanical transport

These defects originate from the film medium, not from the captured scene.

2.2. Optical Characteristics of Film‑Surface Defects

Physical defects share several unique properties:

  • Relief / Micro‑shadowing
    Because dust and scratches sit on top of the film, they create tiny shadows or highlight edges depending on the telecine light path.

  • Non‑photographic contrast
    Their luminance does not match the lighting of the scene.

  • Temporal inconsistency

    • Some defects persist across many frames (scratches).

    • Others appear only for 1–3 frames (dust).

    • Their motion does not follow the scene’s motion vectors.

  • No parallax or depth integration
    They do not belong to the scene’s depth structure.

2.3. Why Current Models Cannot Handle This

Topaz models are optimized for:

  • noise reduction,

  • deblocking,

  • detail enhancement,

  • super‑resolution.

But they are not trained to:

  • detect micro‑shadows,

  • identify non‑photographic artifacts,

  • distinguish surface defects from real scene elements,

  • reconstruct missing image data beneath a scratch.

Thus, they may soften the defect but cannot remove it structurally.

3. Proposal: New AI Model “FilmClean‑Net” (Physical Film Defect Restoration)

3.1. Objective

Develop a model capable of:

  • detecting dust, scratches, fibers, and surface debris,

  • differentiating them from real scene content,

  • reconstructing the underlying image using temporal and spatial cues,

  • preserving film grain and texture.

3.2. Technical Approach

A. Relief‑Based Detection

Use micro‑shadow analysis to identify defects that sit above the emulsion plane.

B. Temporal Consistency Analysis

  • Scratches: long, continuous artifacts across many frames.

  • Dust: short‑lived, inconsistent artifacts.

  • Real objects: follow scene motion, not film transport.

C. Multi‑Frame Reconstruction

Use adjacent frames to rebuild the missing image under the defect.

D. Grain‑Preserving Inpainting

Reconstruct the underlying texture while maintaining:

  • original grain structure,

  • exposure characteristics,

  • film stock identity.

3.3. Implementation Options

Option A: Dedicated Model

Name proposal: FilmClean‑Net
Placed before super‑resolution.

Option B: Pre‑Inference Step

A checkbox in Topaz Video AI:

“Remove physical film defects (dust, scratches, surface debris)”

4. Use Cases

  • Restoration of 8mm / Super 8 home movies

  • 16mm and 35mm archival film scans

  • Telecine transfers with gate dust

  • Film reels with mechanical scratches

  • Historical footage digitized decades ago

  • Professional film restoration workflows

5. Recommended Processing Pipeline

  1. Physical defect detection
    Relief analysis, temporal scanning.

  2. Defect classification
    Dust vs scratch vs fiber vs chemical stain.

  3. Reconstruction
    Multi‑frame synthesis of missing image data.

  4. Grain preservation
    Maintain original film texture.

  5. Super‑resolution / enhancement
    Proteus, Iris, or other models.

6. Why Topaz Should Implement This

  • No consumer‑level tool offers AI‑based relief detection of film defects.

  • Huge demand from film archivists, collectors, and restoration studios.

  • Perfect complement to Topaz’s existing enhancement models.

  • Enables Topaz to enter the professional film restoration market.

  • Dramatically improves the quality of film‑based content before upscaling.

7. Conclusion

A dedicated model for detecting and removing physical film defects — based on relief, micro‑shadowing, and temporal analysis — would fill a major gap in the restoration pipeline.

This model would allow Topaz Video AI to:

  • clean film scans with unprecedented accuracy,

  • preserve authentic grain and texture,

  • differentiate real scene content from surface artifacts,

  • deliver pristine results for archival and cinematic material.

It is a natural and highly valuable evolution for Topaz Video AI.

Best regards, Vincent.