AI Image Compliance Engine

The Challenge

The business problem looked simple at first glance: verify whether a brand mark is visible in a submitted image. In practice, however, the operational requirement was more specific and more demanding. A logo could appear on a shirt, on a separate paper card, in a hand-held placard, on a background object, or as partially readable text. A false positive in any of those cases would weaken the credibility of the compliance process.

• A visible logo is not the same as a compliant logo.
• The system had to reason about where the logo appears relative to the person and the garment.
• The output had to be explainable enough for audit, review, and downstream reporting.

Design Objectives

• Detect the target brand mark reliably across real-world user-provided images.
• Validate whether the mark is inside the shirt region rather than elsewhere in the frame.
• Reduce false positives caused by hand-held paper cards, placards, or visually similar props.
• Generate operator-friendly outputs that can be reviewed, stored, and exported in batch workflows.
• Keep the solution modular enough to extend into APIs, dashboards, or additional compliance classes.

Solution Overview

WeBuildTech approached the problem as a context-aware verification task rather than a standard logo-detection task. The engine first standardises the input image, then uses cloud vision to detect people, faces, logos, and text. From there, it infers the most relevant shirt region, checks whether the logo falls inside that region, and applies a second layer of guards to reject non-garment artefacts such as rigid rectangular paper cards or hand-overlapping signs.

What the system does

Why This Engine Is Different

Most off-the-shelf image classifiers answer a broad question such as whether the brand exists somewhere in the scene. This project required a narrower and more operationally useful answer: whether the brand is on the person's shirt. That distinction is what makes the solution meaningful for compliance workflows.

False-positive control

• Hand overlap checks help identify signs or cards being held in front of the torso.
• Paper-like texture and rectangular-shape cues reduce the chance of mistaking printed placards for garment branding.
• Colour and edge-transition guards compare the suspected logo patch against the surrounding shirt surface.
• The system short-circuits when it reaches a clear answer, balancing quality with API-call efficiency.

Engineering Footprint

Operating Model and Deployment Path

• Current operating shape: local batch CLI and desktop GUI suitable for controlled review workflows.
• Accepted input modes: local files, folders, web URLs, and cloud storage URIs.
• Current outputs: annotated images, JSON reports, and Excel summaries.
• Scalable next step: package the same engine behind an API and connect it to cloud storage, dashboards, and review queues.

Business Value Delivered

• Turns a manual visual check into an automated first-pass compliance decision.
• Produces more defensible outputs by attaching evidence rather than returning a black-box yes or no.
• Improves scalability for branding or field-audit workflows that process many images in batch.