💻 computer science

KnifeHunter: Structured Local Representation Learning for Fine-Grained Knife Image Retrieval in Law Enforcement

This paper introduces KnifeHunter, a practical forensic retrieval system featuring the CoRe-Net architecture and a large-scale dataset, which enables scalable, high-accuracy fine-grained knife identification and has been successfully deployed by UK law enforcement for operational intelligence and source attribution.

Original authors: Syed Sameed Husain, Eng-Jon Ong, Stephen Simpson, Trevor Hamshere, Matt Turner, Miroslaw Bober

Published 2026-08-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Syed Sameed Husain, Eng-Jon Ong, Stephen Simpson, Trevor Hamshere, Matt Turner, Miroslaw Bober

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a specific needle in a haystack, but the haystack is made of other needles, and the one you're looking for has a tiny, unique scratch on its side that no one else has. This is the world of computer vision, a branch of artificial intelligence where computers learn to "see" and understand images. Usually, these systems are great at spotting big things, like "that's a dog" or "that's a car." But when you need to tell the difference between two almost identical twins—say, two knives that look exactly the same except for a tiny pattern on the handle—the computer gets confused. This is called fine-grained retrieval. It's a huge challenge for police and investigators who need to match a knife found at a crime scene to a specific model in a massive catalog to figure out where it came from. If they can't do this quickly, it slows down investigations and makes it harder to stop crime.

Enter KnifeHunter, a new digital detective tool built to solve this exact problem. The researchers behind it realized that looking at a knife photo is tricky because crime scene photos are messy: they might be dark, blurry, have rulers or bags in the background, or show only part of the knife. To teach their computer how to handle this, they built a massive library called the KnifeHunter dataset. It's like a giant photo album containing over 25,000 pictures of 543 different types of knives, taken from police evidence lockers, online shops, and border seizures. They even added over a million "decoy" images—pictures of other objects—to make sure their system doesn't get tricked by random clutter.

The star of the show is a new AI brain they invented called CoRe-Net. Think of most image-recognition systems as a person who looks at a knife from far away and says, "It looks like a kitchen knife." CoRe-Net is different; it's like a detective with a magnifying glass and a wide-angle lens. It uses a special trick called Structured Complementary Representation Learning (SCRL) to break the image down into tiny, specific clues (like the shape of the blade's edge or a maker's mark) and organizes them into a "team" of experts. Then, it uses a Bi-Directional Reciprocal Fusion (BDRF) system to make sure the "wide-angle" view and the "magnifying glass" view talk to each other, correcting each other's mistakes.

The results are impressive. When tested on difficult photos with messy backgrounds, CoRe-Net correctly identified the right knife in the top 10 results about 83.8% of the time, beating all previous methods. Even better, when UK police actually used the system in the field during real operations (called Operation Sceptre) between 2023 and 2025, it was right the very first time 99.2% of the time. This isn't just a lab experiment; it's a tool that is currently helping officers turn a simple photo of a seized weapon into a searchable clue, helping them track down sources and solve crimes faster. The paper suggests that by automating this matching process, they can reduce the heavy manual work police usually have to do, turning hours of visual comparison into seconds of digital searching.

Technical Summary: KnifeHunter and CoRe-Net

Problem Statement
Knife-enabled violence presents a significant public safety challenge, particularly in the United Kingdom. Law enforcement agencies require scalable tools for catalogue-level knife identification, intelligence analysis, and source attribution. Current workflows rely heavily on manual visual comparison and unstructured records, which are specialist-dependent, time-consuming, and difficult to scale. From a computer vision perspective, knife identification is a demanding fine-grained instance-level retrieval problem. Unlike standard landmark retrieval, forensic knife imagery involves subtle inter-class differences (e.g., blade geometry, serration patterns, maker marks) and substantial intra-class variation due to wear, damage, and modification. Furthermore, operational imagery is characterized by uncontrolled illumination, specular reflections, occlusion, background clutter, and the presence of evidence markers (e.g., police rulers), making catalogue-level matching difficult. Existing descriptor aggregation methods often struggle to preserve structured local evidence under these adverse conditions or fail to explicitly organize local and global representations into complementary components within a compact embedding.

Methodology: CoRe-Net
To address these challenges, the authors propose CoRe-Net (Complementary Representation Network), a compact single-descriptor retrieval architecture designed for forensic imagery. CoRe-Net integrates complementary global and local evidence through structured descriptor aggregation and bidirectional fusion. The architecture consists of three primary components:

Weibull-based Global Representation: To mitigate the influence of background clutter and high-magnitude outliers (such as specular highlights), CoRe-Net employs a learnable Weibull-based activation shaping function before global average pooling. This function combines polynomial growth with exponential decay to suppress weak background responses while preventing a few strong activations from dominating the descriptor. The output is followed by power normalization and $\ell_2$ normalization.
Structured Complementary Representation Learning (SCRL): This local aggregation branch organizes spatially localized discriminative evidence into multiple complementary prototype-based representations.
- Saliency-Guided Feature Refinement (SGFR): A variance-based mask is computed to attenuate locally homogeneous regions (e.g., uniform backgrounds) and emphasize high-variance structures (e.g., serrations, grind lines).
- Prototype-Based Local Descriptor Learning (PBLD): Refined features are mapped to a set of learnable prototypes using scaled dot-product attention. This allows tokens to contribute to multiple prototypes with soft weights, organizing local features into a compact latent basis.
- Orthogonality Regularization: An orthogonality loss is applied to the prototypes to prevent collapse and encourage diversity, ensuring different prototypes encode distinct local patterns.
- Prototype Gated Aggregation: A learned mixture of prototype descriptors is computed via a gated mechanism to form a compact local summary.
Bi-Directional Reciprocal Fusion (BDRF): This module couples the global and local branches to enforce complementarity.
- Global-to-Local Residual Projection: The global descriptor is projected out of the refined local feature map to remove components already explained by the global context, forcing the local branch to focus on residual evidence.
- Local-to-Global Reciprocal Injection: The aggregated prototype descriptor is injected into the global descriptor via a learnable gated residual addition. This allows weak but class-separating local cues to refine the robust global representation.
- The fused representation is globally pooled and projected into a 512-dimensional embedding for cosine-similarity retrieval.

The system is trained end-to-end using a sub-center ArcFace classification objective combined with the prototype orthogonality regularizer.

Key Contributions
The paper introduces three primary contributions:

KnifeHunter Dataset: A forensic knife retrieval benchmark comprising 25,843 images across 543 knife classes. The data is sourced from police evidence repositories, retail catalogues, and border-force seizures. It includes structured metadata and defines "Medium" and "Hard" evaluation protocols, as well as a large-scale distractor evaluation set containing over 1.1 million images (including disjoint police knife images and general COCO images).
CoRe-Net Architecture: A novel retrieval architecture that introduces Structured Complementary Representation Learning (SCRL) to organize local evidence into complementary prototypes and Bi-Directional Reciprocal Fusion (BDRF) to integrate global and local evidence. This design supports efficient single-descriptor retrieval without the need for query expansion, re-ranking, or geometric verification.
Benchmarking and Operational Validation: The authors evaluate modern backbones and state-of-the-art aggregation methods under various protocols. Crucially, the system was deployed by UK police forces during Operation Sceptre (2023–2025), providing real-world validation of its efficacy.

Results
Evaluated on the KnifeHunter benchmark using an EVA02-Base backbone, CoRe-Net demonstrates superior performance compared to state-of-the-art baselines (including GeM, DOLG, DELG, TokenNet, SuperGlobal, and SENet).

Medium Protocol (No Distractors): CoRe-Net achieves 88.0% mAP and 86.7% mP@10, outperforming the best baseline (SENet) by 3.1 mAP points.
Hard Protocol (No Distractors): CoRe-Net achieves 70.2% mAP and 67.9% mP@10.
Distractor Conditions (+desc): Under the large-scale distractor protocol, CoRe-Net achieves 85.1% mAP and 83.8% mP@10 on the Medium protocol, and 64.9% mAP and 61.5% mP@10 on the Hard protocol.
Operational Deployment: During field queries in Operation Sceptre deployments, the system achieved 99.2% mP@1, demonstrating reliable top-ranked retrieval in frontline policing workflows.

Ablation studies confirm that each component (Weibull shaping, SGFR, prototypes, and BDRF) contributes significantly to performance, with the prototype-based modeling and reciprocal fusion being particularly critical for handling fine-grained variations and clutter.

Significance and Claims
The paper claims that KnifeHunter provides a practical and effective multimedia retrieval framework for fine-grained forensic knife matching in operational law-enforcement settings. By combining a large-scale, structured dataset with a specialized retrieval architecture, the work addresses the limitations of manual inspection and existing computer vision methods in handling the specific challenges of forensic imagery (clutter, occlusion, subtle differences).

The significance extends beyond technical retrieval metrics to operational impact. The system has been integrated into the workflows of 22 UK police regions, enabling officers to transform routine knife encounters into searchable intelligence. This automation supports structured intelligence gathering on weapon typology and retail availability, contributing to policy reviews and evidence generation for enforcement actions against non-compliant sellers. The authors emphasize that the system operates as a deterministic, CPU-compatible tool suitable for large-scale deployment without requiring complex re-ranking or geometric verification steps.

Technical Summary: KnifeHunter and CoRe-Net

More like this