Automated Adhesive Injection Localization in Aerospace Composites via CLIP-Guided Multimodal Image Enhancement and Adaptive Contour Reconstruction

As a critical component of satellite manufacturing, improving the automation level in composite structural plate production is essential for satellite mass production. We propose a dual-decoder image enhancement framework integrating CLIP (Contrastive Language-Image Pre-training) and VGG architectures. Through illumination correction and geometric deformation augmentation applied to the original dataset, we generate optimized training data. Furthermore, we design an irregular-edged image object contour reconstruction method to identify the center positions of adhesive injection holes during actual filler dispensing. Preliminary experiments validate that the image enhancement method effectively improves image quality based on multimodal features (illumination, deformation magnitude), providing a viable dataset augmentation approach for future YOLO (You Only Look Once) model training. Combined with a preliminarily trained YOLO recognition model, the contour reconstruction method can accurately locate adhesive injection hole positions.