Abstract: In order to address the detection difficulty of small parasite eggs, occluded targets and low resolution in microscopic images, a lightweight object detection model named LCN-YOLOv8(Lightweight Convolution Network YOLOv8) was proposed in this paper. Firstly, a new feature extraction module PA-C2f(Patial Attention C2f) was designed to replace all C2f modules. It incorporated partial convolution with a separated and enhancement attention module, which could reduce the number of parameters and maintain efficient feature extraction ability. Then, Space-to-depth Convolution was introduced to replace the standard convolution, reducing model complexity and feature loss of low-resolution parasite eggs. Finally, a Lightweight High-level Screening-feature Feature Pyramid Network was designed to enhance the model's ability of multi-scale feature fusion and detection performance for occluded parasite eggs. Compared with the baseline network YOLOv8s, LCN-YOLOv8 improves mAP (mean Average Precision)@0.5:0.95 on single and multi-eggs datasets by 1.7 and 1.6 percentage points, respectively, reduces parameters and computation by 6.0×10^6 and 7.5 GFLOPs. Experimental results demonstrate that LCN-YOLOv8 achieves an improved balance between detection accuracy and model complexity, which can assist in the diagnosis of parasitic diseases.

Key words: object detection, parasite eggs, partial convolution, feature fusion, attention module

摘要： 针对寄生虫卵显微图像虫体较小、背景遮挡、分辨率低等检测难题，提出一种基于YOLOv8s的轻量目标检测模型LCN-YOLOv8(Lightweight Convolution Network YOLOv8)。首先，融合部分卷积与分离增强注意力模块，设计特征提取模块PA-C2f(Patial Attention C2f)替换模型中全部C2f模块，减少参数量的同时保持高效特征提取能力；然后，引入空间深度转化卷积代替常规卷积降低模型复杂度，减少低分辨率虫卵特征的丢失；最后，设计轻量高级特征筛选金字塔网络作为颈部网络，加强模型的多尺度特征融合与检测被遮挡虫卵的能力。与基线网络YOLOv8s相比，所提出的LCN-YOLOv8模型在单虫卵与多虫卵数据集上的mAP(mean Average Precision)@0.5:0.95分别提高1.7个与1.6个百分点，模型参数量减少6.0×10^6，浮点运算量下降7.5GFLOPs。实验结果表明，LCN-YOLOv8在检测精度与模型复杂度之间取得较好的平衡，能够辅助诊断寄生虫疾病。

关键词: 目标检测, 寄生虫卵, 部分卷积, 特征融合, 注意力模块