Industry Context: Enhancing Quality Inspection Systems with AI-Powered Computer Vision

In canned food production, surface-level defects — like dents, rust, scratches, or sealing issues — pose serious risks to product quality and brand integrity.

Many manufacturers already use semi - or fully - automated visual inspection systems built by robotics providers or system integrators. However, these systems often rely on rule-based or conventional vision techniques that:

• Struggle to adapt to product variations or new defect types

• Require time-consuming reprogramming for changes

• Can’t deliver the accuracy or learning capability modern factories demand

System integrators and robotics providers building these solutions need smarter, faster ways to embed high-performing, flexible and without the complexity of building AI from scratch.

The Opportunity is adding AI Computer Vision

By integrating AI-powered Computer Vision into these inspection systems, solution providers can:

1. Better detection accuracy across diverse product lines

2. Handle complex defect types and product variations

3. Adaptability to new defects without rewriting code

4. Reduce false positives/negatives compared to rule-based systems

5. Enhanced ROI for end-manufacturers

The Solution: Accelerate AI Computer Vision Development Process With Tuba.AI Work Flow

To showcase Tuba.AI’s impact, we built an end-to-end defect detection pipeline using a public Canned Food Surface Defect Dataset and deployed it entirely through Tuba Workflow Builder — no coding required.

Key Steps using Workflow on Tuba.AI

1. Dataset Labeling:

• Uploaded dataset of canned products

• Used Tuba’s labeling tool for bounding box detection

• Defined two classes: Defected and Non-Defected

2. Model Training:

• Dragged & configured the Train Model block

• Selecting training type “Detection” the same as labeling tool and selecting "Automatic" to enable automatic training. Alternatively, you can select manual operation and select the required parameters.

  • Accuracy vs Time: define what the training process considers most during training.

  • Framework Type: The framework to use. Pytorch is the standard and the go-to choice because of its wide support.

  • OpenVino/TensorRT format: When selected in Framework, choose to what degree quantization should be applied to the weights of the model; float 16, float 32, integer 8

  • Accuracy: which model to choose, the one with the best accuracy or the one with the optimum accuracy, choosing optimum will ensure that the model generalizes better.

  • Time: The maximum time to take during training.

• Training completed with built-in performance tracking

  
  

3. Deployment:

After training, the model became ready to be deployed on cloud infrastructure or edge devices for real-time performance allowing real-time inspection during production.

4. Inference and Performance Evaluation:

• Tested model with new samples

• Visualized predictions with confidence scores and bounding boxes

  
  

5. Integration and Scalability:

This defect detection workflow built with Tuba.AI can seamlessly integrate into automated quality control systems that can be expanded to handle more product variations by updating the dataset and retraining the model.

Results

During model training on the canned food defect dataset:

• Initial mAP (mean Average Precision): 26% — indicating a low starting point typical before optimization

• Peak Training mAP: 99% — the model learned the patterns of defects very well during training

• Final Validation mAP: 91% — a strong indicator that the model generalized well to unseen data without overfitting
  
  

This means the model not only performed well on training images but also maintained high accuracy on test samples — a key measure of real-world reliability.

In visual tests, the model successfully detected multiple types of surface defects  with high confidence scores and precise localization.

This confirms that Tuba.AI’s Workflow Builder can produce robust, production-ready models that adapt to real variation in product appearance, lighting, and defect types.

Business Impact for System Integrators & Robotics Providers

With Tuba.AI, You Get:

• AI Without the Overhead: No need to hire or train AI specialists

• Faster Delivery: Deploy defect detection in days, not months

• Flexible Deployment & Integration: Edge devices, Cloud-ready, On-premise

• Continuous improvement: Retrain on production data for even better accuracy

Tuba.AI enables you to build intelligent defect detection pipelines that are faster to deploy, easier to maintain, and better performing, helping you deliver more value to your clients while reducing development time and cost.

Let’s talk about how Tuba.AI can power your next inspection solution. Talk to our expert now!

Resources 

• Dataset Link : Canned Food Surface Defect Dataset