Combining DETR Multi-scale Perception Structure to Improve Target Boundary Extraction Accuracy in Image Data

Abstract

Accurate target boundary extraction is essential for high-precision image interpretation in intelligent sensing systems and provides important technical support for electromagnetic imaging, remote sensing, and vision-assisted signal perception applications. Owing to the limited feature representation capability of the original Detection Transformer (DETR) when processing objects at different scales, target boundary extraction accuracy remains insufficient, particularly in scenarios involving fine-grained contour localization. To address this issue, an improved DETR framework based on a multi-scale perception structure is proposed. A multi-scale encoding architecture integrated with a feature pyramid network is employed to capture features at multiple resolutions, enhancing boundary-aware spatial representations through an edge completion mechanism. A scale-aware multi-head attention module is incorporated into the encoder to preserve scale consistency during global feature modeling and reduce semantic drift. Furthermore, a boundary regression head combined with positional embedding jointly exploits spatial and semantic information to improve contour localization, while an IoU-aware loss function optimizes prediction accuracy in overlapping regions. Experimental results demonstrate that the proposed method increases the average IoU from 0.81 to 0.86 and improves target boundary AP from 44.9% to 48.1%. Compared with Deformable-DETR and Elastic-DETR, it also achieves lower boundary error ratios, confirming that the proposed multi-scale perception mechanism effectively enhances high-precision boundary fitting and offers practical value for intelligent visual sensing systems in engineering applications.