Improving the Boundary Recognition Performance of Visual Elements on Product Packaging by Integrating the Image Segmentation Mechanism of SAM-Adapter

Accurate boundary recognition under high-noise, complex-texture, and reflective conditions remains a fundamental challenge for intelligent visual perception systems and is equally important for electromagnetic imaging and radar-based target analysis that require reliable feature localization. This study presents a boundary-aware segmentation framework by integrating a gradient-guided Adapter mechanism into the Segment Anything Model (SAM). Pixel-level boundary response maps generated by the Laplacian of Gaussian operator are combined with gradient magnitude and directional sensitivity weights to guide adaptive feature calibration, while boundary-aware cross-entropy loss and Hausdorff distance constraints jointly optimize parameter learning and geometric consistency. The proposed strategy enhances multi-scale boundary representation without sacrificing the lightweight characteristics of Adapter-based fine-tuning. Experimental results demonstrate that the method achieves a Boundary Intersection over Union (BIoU) of 0.923 under 20% noise interference and maintains a contour F1 score of 0.951 under 20% reflectivity, significantly outperforming existing comparison methods in boundary continuity and localization accuracy. The framework provides an effective solution for high-precision image segmentation and offers potential value for electromagnetic imaging, microwave sensing, and intelligent perception systems requiring robust boundary extraction under complex interference conditions.