1.1 Basic information
The ML model takes an image as an input, and returns the locations, confidence scores, and classifications of where it believes defects are present in the current image. More information on each of the outputs:
- x1, y1, x2, y2: the coordinates of the box where the defect is within the frame.
- confidence score: this is a score between the values of 0.0 – 1.0. It represents how certain the ML model is of its prediction. This output may be misleading, as on some printers with their lighting conditions, camera qualities, orientations, etc, will detect defects at very high values (example = 0.95), and on other printers with different conditions will detect with medium values (example = 0.55). For this reason it is important to Configure your Setup.
- classification: this is a value anywhere between 0.0 and N, where N is the amount of defect types supported at the current time by PrintWatch. Each positive integer represents a type of defect.
1.1.1 False Positives
A False Positive (FP) is when the ML model detects a region in the image as having a defect. This can occur for a variety of reasons, some of them being:
- Poor lighting, especially overexposure can lead to regions being misidentified as defects. Ensure that the print bed is well-lit and not overexposed. Extruder lights (A light on the extruder that only lights up the area that the print head is near) are not recommended for accurate AI detections.
- Items in the background can throw off the AI. Some real-world items may vaguely resemble defects, and the ML model will detect those items as defects. Ensure that the background in the image is as static as possible (the best is a solid colored wall or enclosure). Additionally, noise from the camera may reduce the ML models ability to detect. Some cameras quality may come out as grainy, with bad color/bith depth, or simply low quality, and this will hinder the ML models ability to detect. Ensure you are using a camera with the minimum resolution of 720p, and preferably 1080p (some 720p cameras commonly have grainy quality).
- Leftover filament/scraps near the print bed:
- Leftover filament and scraps that resemble defects will be identified as defects. Although the ML model is acting as expected, this is not desirable. Ensure that the print bed and surroundings are clear of any debris.
1.1.2 Not detecting True Positives
Although the Machine Learning model is very accurate, it may miss some defects, or detect them with a lower than expected confidence score. It is important to have a good setup as dictated in the Configuring your Setup document. Some setups may detect defects at lower confidence than others, and this is expected behavior. If your setup is setup correctly and you are experiences issues with detection, reach out to support or contact us via the discord channel.