We propose a two stage detection system for identifying surface deformities such as dents and dings. The first stage uses structured light scanning to capture detailed geometric data and detect deviations from the expected surface profile. In the second stage, a machine learning classification pipeline filters out false positives such as screws, rivets, and seams by distinguishing true deformities from non-defective features. This approach enhances the overall accuracy and reliability of automated inspection processes.<\/p>\n\n\n
\n<\/figure>\n<\/div>\n\n\n
Pipeline of our two-stage detection system<\/p>\n\n\n\n
First Stage Detector<\/h2>\n\n\n\n
In the first stage of the pipeline, a laser line is projected onto the surface of the plane to capture signals that may reveal subtle dents, which are often difficult to detect visually. The deformation of the laser line serves as an indicator of potential dents. After capturing the laser-projected image, we calculate the average projection distance of each pixel on the laser line relative to the main line direction, determined via PCA. Since the laser line may be distorted by surface dents or other artifacts, it typically deviates from a perfectly straight path. We then plot these projection distances as a curve and heuristically identify local minima as candidate dent locations on the plane.<\/p>\n\n\n
\n<\/figure>\n<\/div>\n\n\n
Illustration of deformation curve<\/p>\n\n\n\n
Two Camera Stereo Setup<\/h2>\n\n\n\n
In practice, the deformation of the laser line is often minimal (approximately 1\u20132 pixels). To improve the accuracy of the first stage, we require a very clean laser-line image. However, the second-stage ML classifier still needs the full scene with background information. To satisfy both requirements, we capture images using two cameras: one with low exposure and one with normal exposure. The low-exposure camera produces an image in which only the laser line is visible\u2014ideal for the first-stage pipeline\u2014while the normal-exposure image is used as input for the second-stage ML classifier.<\/p>\n\n\n\n<\/figure>\n\n\n\n
Two Camera Stereo Setup: low-exposure and normal-exposure<\/p>\n\n\n\n
Second Stage Deformity Classifier<\/h2>\n\n\n\n
After detecting deformities, we extract pixel patches around the identified regions in the normal exposure image and pass them through a machine learning classifier, such as YOLOv11, to categorize them as either Dents and Dings or False Positives. The False Positive category includes features like rivets, seams, and screws, which are not considered true surface deformities. This classification step helps refine the results by eliminating irrelevant detections.<\/p>\n\n\n\n<\/figure>\n\n\n\n
Second Stage Pipeline to remove False Positives<\/p>\n\n\n\n
Pipeline We propose a two stage detection system for identifying surface deformities such as dents and dings. The first stage uses structured light scanning to capture detailed geometric data and detect deviations from the expected surface profile. In the second stage, a machine learning classification pipeline filters out false positives such as screws, rivets, and … <\/p>\n
![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team5\/wp-content\/uploads\/sites\/132\/2025\/12\/Pipeline_full_updated-1-1024x724.png\")
<\/figure>\n<\/div>\n\n\n![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team5\/wp-content\/uploads\/sites\/132\/2025\/05\/deformation_curve-1024x430.png\")
<\/figure>\n<\/div>\n\n\n![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team5\/wp-content\/uploads\/sites\/132\/2025\/12\/image-4-1024x447.png\")
<\/figure>\n\n\n\n![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team5\/wp-content\/uploads\/sites\/132\/2025\/12\/image-5-1024x281.png\")
<\/figure>\n\n\n\n