PCB Defect Detector

Inspiration

Some of us are electrical engineers and thought if we had this in our work environments, it would increase efficiency in working with the PCB's.

What it does

As it stands it detects 6 of the most common types of errors in PCB's. These defects are usually easy to miss since they are so small. You can take a picture or upload a picture from your device and it instantly scans it to determine the defects. It will not only identify where the defects are, but what type of defect it is. You can adjust the confidence interval of the program to identify errors with greater accuracy.

How we built it

We decided to use python for our backend and a html, css, javascript stack for our frontend. The backend used a yolov8 dataset which contains over 8000 images that we used to train our model. We switched to kaggle for a smaller dataset that would actually train our model in time for submission, however, the logic used for yolov8 is still present.

Challenges we ran into

Since yolov8 is such a large dataset, we were having many issues training our model. We attempted to use different datasets that had less content, finally achieving success with the kaggle dataset. We got our model fully trained using 50 epochs. The submission into github was far too large with this data, so it had to be removed for the submission.

Accomplishments that we're proud of

The thing that made us the most proud was how we were able to make an AI that can actually detect errors in things we use in our work environments. Thinking about the applications and how it will make working with PCB's easier for electrical engineers everywhere is very exciting.

What we learned

We learned mainly how we can go about training a model. This can be applied to many different projects that we work on in the future.

What's next for PCB Defect Detector

We plan to add more defect types that way it captures even more defects and add even more reference images to train it on, making it far more accurate. The end goal would be to hopefully make this usable for everyone.