Forensics, visual search & legal tech
Law enforcement & IP protection
ImageTwin — image similarity analysis
Measured results
−90%
manual image search time
Source: Epsilon AI Success Stories & Project Highlights, Ver. 3 — 15 October 2025, page 13
The problem
Comparing an image against a large database was a manual, sequential task, which put a hard ceiling on how much could be checked.
What was built
Vector-based similarity matching that returns visually similar images instantly, including from degraded originals.
Technical approach
- Visual embeddings
- CLIP and ViT embeddings representing each image as a vector.
- Image enhancement
- Preprocessing that recovers usable signal from occlusion, blur and low resolution.
- Nearest neighbour search
- FAISS and Annoy indexes for high-speed similarity lookup.
Measurement framework
Deployment evidence is drawn from the published engagement. The delivery and measurement plans are case-specific recommendations grounded in the documented solution and results.
- Evidence status
- The source documents a working visual-similarity search capability, supported by a reported reduction in manual image-search time.
- Indicative delivery plan
- Prepare and govern the image corpus, generate and index embeddings, benchmark retrieval on known matches, run a blinded analyst pilot, then scale the approved index.
- Recommended baseline
- Recommended: analyst minutes per search, number of candidates manually reviewed and verified-match rate on a fixed pre-deployment query set.
- Recommended measurement method
- Benchmark the same query images against the same frozen corpus before and after vector indexing. Measure system latency, human review time and precision/recall at a fixed result depth, using blinded analyst verification.
Recommended items are not presented as the original study method. They should be replaced with the confirmed project methodology when the underlying records are approved for publication.