Acceptability and challenges of community-based rehabilitation (CBR) of leprosy patients with grade-2 disability in Bangladesh.

End-to-end traceability of surgical instruments is fundamental to infection control and patient safety in modern healthcare systems []. Across decontamination, cleaning, packaging, sterilization, distribution and intraoperative use, each instrument must be uniquely identified and its handling history recorded to support quality assurance, error prevention and adverse-event management []. To enable management at the individual instrument level and full-process traceability in both regulatory and clinical practice, permanent direct part marking is typically applied to the instrument itself, most commonly in the form of engraved alphanumeric strings or two-dimensional codes linked to traceability records []. In real-world settings, however, reliable localization of these markings is hindered by specular reflection from metallic surfaces, highlight saturation, motion blur and scale variation. These effects are particularly detrimental to elongated characters, whose fine strokes and edge cues are prone to fragmentation or disappearance, thereby compromising stable detection and precise localization. In addition, the efficiency of code reading and the operator’s level of proficiency directly influence the cost and practicality of traceability workflows at the individual instrument level []. Developing robust methods for the detection and localization of alphanumeric codes on surgical instruments under challenging illumination and dynamic imaging conditions is therefore of substantial practical importance. Text detection methods for industrial and natural scenes can generally be grouped into two main paradigms: regression-based and segmentation-based approaches []. Regression-based methods typically represent text instances as horizontal boxes, rotated boxes or quadrilaterals, and directly predict their geometry. CTPN, proposed by Tian et al. [], employs fixed-width vertical anchors and combines CNN features with sequence modeling based on RNNs to connect text proposals, achieving strong performance on horizontal text. EAST, introduced by Zhou et al. [], directly regresses text geometry using an end-to-end fully convolutional network, offering high inference efficiency and adaptability to multi-oriented text, but exhibiting limited robustness when dealing with elongated text, curved text and cluttered backgrounds. Segmentation-based methods, by contrast, describe text regions using pixel-wise probability maps and recover final instances through separation or aggregation strategies. PAN [] improves speed substantially while maintaining competitive accuracy through a pixel aggregation mechanism, although there remains room for improvement in scenarios involving extreme aspect ratios. PSENet [] uses progressive scale expansion to separate adjacent text instances, effectively alleviating adhesion between neighboring texts. FCENet [] enhances curved-text modeling through Fourier contour representations. Mask R-CNN [], as a general-purpose instance segmentation framework, provides strong region-level modeling capacity but is often associated with higher computational and deployment costs in text detection tasks involving high resolution images. Among segmentation-based methods, DBNet, proposed by Liao et al. [], integrates the thresholding process into end-to-end optimization through differentiable binarization, achieving a favorable balance between detection accuracy and inference efficiency. It has therefore become a widely adopted baseline in practical applications. Its successor, DBNet++, further improves multi-scale text detection and robustness through adaptive scale fusion and related refinements []. Despite these advances, surgical instrument code detection remains particularly challenging for two main reasons. First, degradations such as reflection, highlight interference and blur destabilize character edges and stroke structures, making elongated engraved codes more susceptible to fragmentation, adhesion and missed detection. Second, many existing text detection networks remain computationally heavy and parameter-intensive, limiting inference efficiency and deployment feasibility. To address these limitations, we propose an improved DBNet-based framework for surgical instrument code detection that simultaneously enhances detection performance and reduces model complexity. The main contributions of this study are as follows.