Enhancing Quality Control with Computer Vision

Computer vision is revolutionizing how computers interact with the visual world. By teaching machines to see and understand images and videos, we're unlocking a world of possibilities. From automating mundane tasks to enabling groundbreaking innovations, computer vision is transforming industries and shaping the future.

In manufacturing, computer vision is enhancing defect detection and reducing costs by identifying problems early to prevent waste. Computer vision, working in conjunction with quality assurance, allows for the collection of data on malfunctioning upstream systems to help pinpoint root causes and reduce the significant financial impact manufacturing defects can have on a business.

The general workflow of a computer vision solution requires 5 steps:

1. Camera Placement.  Valuable product angles can help identify defects, providing a critical service to business owners. Note, this is where you can get creative and use the properties of the product to your advantage.

2. Data Collection. Gathering images of clean and defective products is required to begin training the model. A great start to determining if the solution is feasible is to build a data collection of approximately 100 images.

3. Annotation. Label photos by drawing bounding boxes around the areas of interest, which creates a dataset for your model to learn from.

4. Training. With the annotation complete, the dataset is passed to the model to train. The training process is a hands-off step while the computer processes the data. Depending on the size of your dataset, it may take some time to complete.

5. Testing. With the model trained, all that is left is to feed it images/videos and watch the results.

Proof of Concept: Bolt Defect Detection

Iterating through this process efficiently can help provide quick feedback about solution options. 

As an example, a rapid iteration for determining manufacturing defects in bolts took about an hour when following those five steps.

1. Camera Placement. An iPhone recorded a video of bolts being pulled past the camera. 

2. Data Collection. The video was split into frames, yielding around 300 images

3. Annotation. About 100 images were annotated, with bounding boxes drawn around bolts and defect areas.

4. Training. The model was trained using the dataset to learn about "bolts" and "bad thread," yielding the following results:

5. Testing. Bolts were again pulled in front of the camera to test the newly trained model.

This rapid-proof of concept serves as a baseline for future development. To ensure continued progress, we can start with data from day one to track improvements and work toward a successful implementation. The next steps include enhancing camera quality/focus, refining the dataset, retraining the model, and retesting.

Contact us to see how MichiganLabs can work with you on enhancing defect detection in your business.