Abstract:

In the preprocessing stage of fully automatic medical inspection pipelines， a test tube and liquid level detection algorithm based on improved YOLOv8 was proposed to address the issues of low recognition rate and low speed of test tube sizes and liquid levels in traditional detection methods. Firstly， a lightweight ADown module was employed to replace the Conv module in the backbone network for feature extraction and downsampling， thereby extracting more effective information while reducing the model size. Secondly， Bi-directional Feature Pyramid Network （BiFPN） was utilized for feature fusion， and more hierarchical feature information was integrated through bidirectional and skip connections. Additionally， Omni-dimensional Dynamic Convolution （ODConv） was introduced at the neck， and the C2f-ODConv module was designed to enhance feature extraction capabilities. Finally， the Inner-CIoU bounding box loss function was introduced to accelerate model convergence by auxiliary bounding boxes. Experimental results demonstrate that the proposed algorithm achieves improvements of 3.6， 4.8 and 5.0 percentage points in precision， recall and mean Average Precision （mAP）@50， and a decrease in computational cost （FLOPs） by 13.6% on a self-made dataset. It can be seen that the proposed model can realize accurate recognition of test tubes and liquid levels in real scenarios.

Key words: YOLOv8, test tube detection, liquid level detection, Bi-directional Feature Pyramid Network (BiFPN), Omni-dimensional Dynamic Convolution (ODConv), auxiliary bounding box

摘要：

在全自动医疗检验流水线的前处理阶段，针对传统检测方法检测试管规格和液位的识别率低、速度慢等问题，提出一种基于改进YOLOv8的试管及液位识别算法。首先，采用轻量化ADown模块替换主干网络中的Conv模块以进行特征提取和下采样，从而在减小模型大小的同时提取更多有效信息；其次，采用双向特征金字塔网络（BiFPN）进行特征融合，通过双向连接和跳跃连接融合更多层次的特征信息；同时，在颈部引入全维动态卷积（ODConv）并设计C2f-ODConv模块，以增强特征提取的能力；最后，引入Inner-CIoU边框损失函数，从而通过辅助边框加速模型收敛。实验结果表明，所提算法在自制的数据集上的精确率、召回率和平均精度均值（mAP@50）分别提升了3.6、4.8和5.0个百分点，计算量（FLOPs）下降了13.6%。可见，所提模型可实现真实场景下对试管与液位的准确识别。

关键词: YOLOv8, 试管检测, 液位检测, 双向特征金字塔网络, 全维动态卷积, 辅助边框