Harness machine learning and image processing to deliver accurate brand logo recognition, model training pipelines, and reliable visual identification at scale.
Client Overview
A brand monitoring and analytics company was providing brand exposure tracking services to major consumer goods clients — measuring how often and where their logos appeared across social media posts, online videos, broadcast content, and user-generated imagery. The company's existing approach relied on a combination of keyword-based social listening tools and a small team of human analysts who manually reviewed flagged content to confirm logo presence. This process was slow, could not operate at the scale required by enterprise clients, and was blind to visual content that contained no text reference to the brand. Clients were paying for comprehensive brand exposure data across their entire digital footprint, but the company could only deliver reliable data for content that contained text references to the brand name. Logo appearances in images — a product placed on a table in an influencer's lifestyle post, a branded vehicle in a street photograph, a logo visible in the background of a sporting event — were largely invisible to the existing monitoring approach. This was a significant gap because visual brand placement, particularly in organic and earned content, was exactly the type of exposure that clients most wanted to understand and quantify. As social media volume and video content production continued to grow exponentially, the company recognised that manual review at any meaningful scale was simply not viable. They needed a computer vision solution capable of detecting their clients' logos accurately across diverse, uncontrolled real-world image environments — variable lighting, partial occlusion, different orientations, and diverse backgrounds — and doing so at the processing speed required to handle thousands of assets per day per client.
Our Approach
Zentric Solutions built a custom logo detection pipeline using YOLOv8 as the core object detection architecture. For each client brand, a custom detection model was trained on a curated dataset of positive logo examples sourced from diverse real-world contexts alongside carefully selected negative examples, ensuring the model could distinguish target logos from visually similar competitor marks. Training pipelines were deployed on AWS SageMaker, enabling efficient model iteration and evaluation at scale without requiring on-premises GPU infrastructure. The inference pipeline was designed to process image and video frames through the trained models at high throughput, with confidence-scored bounding box detections returned for every identified logo instance. For video content, frames were extracted at configurable sampling rates and each frame processed through the detection model, with temporal clustering applied to group detections across consecutive frames and deduplicate exposure counts. All detection results — including bounding box coordinates, confidence scores, source asset metadata, and timestamps — were written to a PostgreSQL database for downstream reporting. A REST API layer exposed detection capabilities to the company's existing monitoring platform and client reporting dashboards, allowing detected logo exposures to be surfaced in the analytics interface that clients already used. The company was able to offer genuine image-based brand monitoring for the first time, covering visual logo appearances entirely absent from their previous text-based approach. Detection accuracy across diverse real-world image conditions exceeded 90% for well-represented logo variants, and the system processed volumes that would have required a team of dozens of human analysts working continuously.
Tech Stack
Project Tags
Everything you need to know about this project and our approach.
Zentric Solutions delivers cutting-edge digital products that streamline operations, enhance engagement, and drive lasting growth.
