Application and generalization analysis of orientation-adaptive convolution in high-precision detection of dental burs

Abstract Conventional object detection methods struggle with slender objects exhibiting high aspect ratios, such as dental burs, because the fixed sampling grid of standard convolutional kernels is misaligned with object geometries. This misalignment introduces substantial background noise and severely degrades detection accuracy. To address this issue, we propose an Orientation-Adaptive Convolution (OAC) module. The OAC predicts the object orientation and scale parameters through a lightweight subnetwork, explicitly modelling the aspect ratios and generating rotation-aligned rectangular receptive fields. This enables precise alignment between the convolutional kernels and object geometry. On our self-constructed dental bur detection dataset, OAC significantly outperformed the baseline model, achieving a 96.2% mAP and 93.0% recall, representing improvements of 5.5 percentage points over the baseline. Generalisation experiments have demonstrated that OAC can be flexibly integrated into diverse backbone networks and detection frameworks, consistently improving performance across various architectures and task scenarios, thereby exhibiting excellent generality and scalability. In summary, the OAC effectively addressed the alignment challenge of slender, direction-sensitive targets. It surpasses existing methods in terms of detection accuracy, model efficiency, and generalisation capability, offering a practical solution for the high-precision visual inspection of industrial components with similar characteristics.