Image safety filters
6 months
2 FTE: Head of Computer Vision, ML Engineer
Project
K2 developed a cutting-edge AI application, achieving 99% accuracy in real-time detection of inappropriate content.
Results
The deployed application exhibited superior performance compared to benchmark architectures, including LAION, Gantman, and Azure, as per the client's problem definition. Delivered as a microservice, 99% accuracy affirming the system's efficacy in real-world conditions.
Media Press
In response to the client's imperative to enhance digital safety, K2 Summit undertook the development of an advanced image detection application. The primary objective was to identify images containing erotic, toxic nudity, or violent content, allowing end-users to define these categories based on individual preferences. Notably, the client had previously developed an in-house solution for image detection, but its performance fell short of expectations, leading them to entrust K2 Summit with the task of creating a more effective system.
Delivering sustainable value.
Deliverables
Impact
Skills
• Precision Enhancement: Achieved a 99% accuracy rate in detecting inappropriate content, reducing misclassified records by over 50%. • Rapid Detection: Implemented AI algorithms, specifically deep learning, enabling real-time content analysis at less than 0.5 seconds per photo. • Performance Superiority: Outperformed other architectures (LAION, Gantman, Azure) by 20% based on the client's defined problem parameters. • Adaptable Microservice: Delivered the application as a versatile microservice, seamlessly integrating it into the client's existing systems. • Data Enrichment: Expanded the dataset with over 80,000 images, addressing shortcomings in the client's previous solution.
The deployed application exhibited superior performance compared to benchmark architectures, including LAION, Gantman, and Azure, as per the client's problem definition. Delivered as a microservice, the system seamlessly integrated into the client's environment. Post-implementation analysis showcased a remarkable 99% accuracy, reducing misclassified records by over 50%, and affirming the system's efficacy in real-world conditions.
Image Detection, Deep Learning, Microservices, Digital Safety, Precision Enhancement, Real-Time Analysis