High-Accuracy Grading
The CNN model provides reliable and objective marbling scores, achieving 98% accuracy compared to traditional human evaluation, ensuring consistent quality assessment across all steaks.
Technologies used
Python
PyTorch
PyTorch Lightning
Pandas
OpenCV
React Native
Docker
Motivation
The client, a U.S. food manufacturer, needed to standardize steak quality inspection on its production line. Manual marbling grading was time-consuming and subjective, leading to inconsistent quality assessments. Azati developed a mobile application powered by a CNN model to automate marbling evaluation, providing fast, objective, and repeatable results under production conditions.
Main challenges
Challenge 01
Inconsistent Human Grading
Manual steak grading was highly subjective, with different evaluators often assigning different marbling scores to the same piece of meat. This inconsistency slowed quality inspection on the production line and made it difficult to maintain standardized product quality across production batches.
#1
Challenge 02
Eliminating lighting glare in steak analysis
Bright production-line lighting caused glare on the surface of the steaks, which could be misinterpreted by AI models as fat, leading to inaccurate marbling grades. The system had to incorporate preprocessing techniques and controlled imaging conditions to minimize these errors and ensure reliable analysis.
#2
Challenge 03
Rapid and Reliable Feedback Required
The client needed near-instantaneous grading results to support continuous production-line inspection. Achieving both high accuracy and fast feedback was challenging, as the model had to process images quickly without compromising grading precision, while the mobile app needed to deliver results seamlessly to users on the floor.
#3
Our approach
Understanding Client Needs
Azati conducted a discovery call to understand the client's production workflow, quality inspection process, and grading requirements.
Training a CNN Model for Marbling Detection
Azati trained a CNN using a client-provided dataset, ensuring the model could accurately detect and classify marbling levels across various steak images.
Mobile App Development and Iterative Refinement
The mobile application was designed and developed as an MVP,
allowing quality inspectors to capture or review steak images during production-line inspections. User feedback led to three iterations, further enhancing functionality and user experience.
Enabling Seamless Model Updates
The app was built to allow easy updates of the AI model without requiring a full rebuild, ensuring scalability and maintainability.
Facing the same challenge?
Bring your complexity. We'll bring the plan. Tell us about your project and we'll get back within one business day.
Inquire for more infoSolution
01
CNN-Powered Image Analysis
Leverages a convolutional neural network trained on client-provided datasets to automatically detect and classify steak marbling levels from images. Handles variations in lighting, steak orientation, and fat distribution to ensure accurate grading.
Key capabilities:
- Automatic marbling detection and classification
- Handles diverse lighting and steak orientations
- High accuracy comparable to human graders
02
User-Friendly Mobile Interface
Provides an intuitive mobile application interface for capturing or reviewing steak images during quality inspection and receiving immediate grading results.
Key capabilities:
- Capture and process steak images easily
- Instant feedback on marbling grades
- Minimal training required for staff
03
Flexible AI Model Updates
Supports seamless updates of the CNN model without requiring a full app rebuild, allowing continuous improvement and adaptation to new grading standards or datasets.
Key capabilities:
- Incremental model updates without app redeployment
- Easy integration of new datasets
- Ensures long-term scalability and maintainability
Results & business impact
Reduced Manual Effort
The solution reduces manual effort required for production quality inspection while minimizing grading variability.
Improved Operational Efficiency
Faster grading supports continuous production-line quality control, helping maintain throughput while improving grading consistency.
Flexible and Scalable Solution
The app supports incremental AI model updates and iterative improvements, ensuring the solution remains accurate and adaptable as grading standards or datasets evolve.
Enhanced User Experience
An intuitive mobile interface allows non-technical staff to use the system efficiently, with instant feedback and seamless integration into daily operations.
