Car Paint Defect Detection Based on PMD With Multisource Image Fusion

In order to achieve automatic detection of car paint defects, this study proposes an algorithm based on phase measuring deflectometry (PMD) with multisource image fusion. First, a dataset containing 2502 sets of car paint defect images was collected and analyzed based on PMD. Then, a multisource image fusion framework was proposed to improve the detection performance. In this framework, the absolute phase map and four sinusoidal fringe images with different phases are used as the source images to provide richer features. Additionally, a fusion label was designed to guide the fusion network in learning semantic information. Finally, the general object detection framework you only look once version 11 (YOLOv11) is trained on the fused images to locate and identify defects. The experimental results indicate a significant enhancement in paint defect detection performance with PMD, as the metric mAP@0.5 increased from 0.639 to 0.856. The proposed multisource image fusion framework significantly improves the balance of detection accuracy across all defect classes, as the metric class standard deviation reduced from 0.125 to 0.072.