ADR 003: Use Perceptual Hashing for Frame Deduplication

Status

Accepted

Context

Video frames are extracted at configurable rates (e.g., 0.5 fps), resulting in many similar frames. We need to:

• Remove duplicates to reduce storage and processing
• Preserve content changes (slide transitions, scene changes)
• Process efficiently (potentially thousands of frames)

Decision

Use Perceptual Hashing (pHash) with Hamming distance comparison.

Consequences

Positive

• Fast computation (O(n) for n frames)
• Robust to minor changes (compression, slight variations)
• 64-bit fingerprints enable fast comparison
• Tuning via Hamming distance threshold
• No ML model required

Negative

• May miss semantic duplicates (different visuals, same content)
• Not robust to major transformations (crops, rotations)
• Threshold requires tuning per content type

Algorithm Details

import imagehash
from PIL import Image

def phash_dedup(frames, threshold=10):
    unique = []
    last_hash = None
    
    for frame in frames:
        img = Image.open(frame.path)
        current = imagehash.phash(img)
        
        if last_hash is None or (current - last_hash) > threshold:
            unique.append(frame)
            last_hash = current
            
    return unique

Alternatives Considered

Alternative	Pros	Cons
SSIM	More accurate	Much slower
CNN embeddings	Semantic understanding	Requires GPU, model
Pixel diff	Fast	Not robust to encoding
aHash/dHash	Faster	Less accurate
Feature matching	Handles transforms	Complex, slow

Threshold Guidelines

Content Type	Recommended Threshold
Slides/Presentations	8-10
Screen recordings	10-12
Talking head	12-15
Dynamic video	15-20

Notes

Consider running SSIM on pHash-filtered candidates for higher accuracy if needed.