The idea of this project it's to have a machine learning app that will detect and identify bottlecaps from a source.
bottle-caps uses computer vision techniques, primarily based on OpenCV, to identify and match bottle caps. While the app currently operates locally, future plans include transitioning to cloud hosting. The app utilizes key OpenCV methods including Simple Blob Detector, Hough Transform Circles, MobilNet, each requiring specific preprocessing steps.
Below is an overview of how the code operates.
Currently, there is a script that generates all the data and uploads it to Pinecone, the vector database used to compare the vector result from the model and the ones from the database: scripts/generate_model.py.
Initially, the Hough Transform Circles method was used, but it resulted in multiple false positives due to the variability in circle sizes. To resolve this, the Simple Blob Detector was applied first. This process involves reducing the image's color depth to three levels (photos should ideally be taken on a white background or with high-contrast caps). Overlapping blobs, which might represent multiple detections of the same cap, were removed. The median size of the blobs was calculated to ensure accuracy.
- Why Median Over Average? Using the median rather than the average prevents large or small false positives from skewing the results. Since all bottle caps have a diameter of 37mm, the median provides a more reliable measurement.
Based on the approximate radius determined from the blobs, upper and lower bounds were set. Preprocessing was applied before using Hough Transform Circles to identify the positions of all circles in the image.
Currently using a MobilNet model to identify the caps, basically transform them in vector format and then saving them in Pinecone. It is a vector database great for comparing AI content.
In the following images you can see how the detect method works and a small animation using the visual module.
Identification model using only a pretrained MobilNet
Below is an example output of the full model. The results show areas for improvement: while the first two caps are correctly identified, the bottom two are labeled incorrectly. Speed optimizations are also needed. For additional notes on testing, refer to ScriptsMain/note_tests/README.md.
- The code from the visual folder was written with ChatGPT as I did not want to dive too much into the creation of
.gifmyself.