改进YOLO11的学生实验过程行为检测算法

Improved YOLO11 Algorithm for Detecting Student Behavior During Experiments

摘要:

针对学生实验行为分析中，因采集设备性能及环境干扰导致的图像模糊、分辨率低和遮挡等检测难题，提出一种基于改进YOLO11的学生实验过程行为检测算法：HMS-YOLO11，算法引入高频增强残差块（HFERB）改进C3k2模块，增强对图像高频细节（如边缘、纹理）的提取能力，从而提升模型在低分辨率场景下的特征判别能力；设计了多分支辅助特征金字塔网络（MAFPN），该网络通过细节增强分支和语义增强分支，增强多尺度特征融合与空间表征能力，提升检测精度；并在检测头中嵌入轻量级SEAM注意力模块，减少遮挡带来的漏检与误检问题。实验结果表明，HMS-YOLO11改进模型在自建数据集上mAP50达到90.9%，mAP50-95达到66.5%，Recall达到89.1%，对比基准模型分别提高了2.2%、1.7%、4.6%，并且模型参数量和计算量更低。

Abstract:

To address detection challenges such as image blur, low resolution, and occlusions in the analysis of student experimental behaviors often caused by equipment limitations and environmental interference, this paper presents an improved YOLO11-based algorithm named HMS-YOLO11 for detecting student behaviors during experiments. The algorithm introduces a High-frequency Enhanced Residual Block (HFERB) to enhance the C3k2 module, strengthening the extraction of high-frequency details (e.g., edges and textures) and thereby improving feature discriminability in low-resolution scenarios. A Multi-branch Auxiliary Feature Pyramid Network (MAFPN) is designed, which incorporates detail enhancement and semantic enhancement branches to boost multi-scale feature fusion and spatial representation, leading to higher detection accuracy. Furthermore, a lightweight SEAM attention module is embedded into the detection head to effectively reduce missed and false detections caused by occlusions. Experimental results show that the proposed HMS-YOLO11 achieves 90.9% mAP50, 66.5% mAP50-95, and 89.1% recall on a self-built dataset, surpassing the baseline model by 2.2%, 1.7%, and 4.6%, respectively, while also reducing both parameter count and computational cost.