A lightweight alignment-aware DBNet for surgical instrument code detection.

Abstract

Reliable detection of engraved surface codes on surgical instruments is essential for end-to-end traceability, yet remains challenging in practice because metallic reflection, motion blur, scale variation and weak textures often hinder stable localization. Here we present LA-DBNet, a lightweight detection framework built on DBNet for this task. The model uses MobileNetV4 with LiteFPN to reduce complexity while preserving multi-scale feature representations. To better capture the elongated structure and edge features of engraved codes, we introduce a Directional Edge Collaborative Alignment (DECA) module to improve cross-scale feature alignment, and embed an Efficient Channel Attention (ECA) mechanism in the high-resolution feature layer to enhance responses relevant to the target and suppress noise caused by reflections. We further incorporate a region-weighted consistency learning strategy during training to improve robustness to degraded samples. On our surgical instrument code dataset, LA-DBNet achieves an F1 of 95.8%, improving DBNet by 3.6 percentage points, while reducing parameters to 3.35 M and reaching 33.6 FPS. On ICDAR2015, it attains an F1 of 86.1%. These results show that LA-DBNet improves detection performance while substantially reducing model size and maintaining efficient inference in surgical instrument code detection.