基于轻量级网络的小目标检测算法

Small object detection algorithm based on lightweight network

针对YOLOv5算法在检测小目标时存在准确率较低的情况,提出旨在提高小目标检测准确率的轻量级网络KOS-YOLOv5算法。首先采用K-means++聚类技术选择一组合适的锚框尺寸作为模型的先验,对小目标实现更精确的锚框尺寸,使模型能适应不同大小的目标;其次利用简化正负样本分配策略(SimOTA)进行动态样本匹配,更好地优化损失函数;最后将空间上下文金字塔(SCP)模块集成到算法检测层中,促使骨干网络更加关注小目标的特征信息,用以增加目标特征提取能力,提高目标的检测准确率。结果表明,改进后的KOS-YOLOv5算法与传统的YOLOv5模型进行比较,算法在检测精确度(P)方面提高了4%,召回率(R)方面提高了2.4%,平均检测精度(mAP)提高了3.1%,损失函数值(Loss)降低了5%,最终检测精度为95.38%。

In view of the low accuracy of YOLOv5 algorithm in detecting small objects,a lightweight network KOS⁃YOLOv5 which aims to improve the accuracy of small object detection is proposed.The K⁃means++clustering technology is used to select a set of appropriate anchor box sizes as priors for the model,achieving more accurate anchor box sizes for small objects and enabling the model to adapt to objects with different sizes.A simplified positive⁃negative sample assignment strategy,which refers to SimOTA,is utilized for dynamic sample matching,so as to optimize the loss function more effectively.The spatial context pyramid(SCP)module is integrated into the algorithm's detection layer,which encourages the backbone network to pay more attention to the feature information of small objects,so as to enhance the capability of extracting object features and improve the accuracy of object detection.The results show that the improved KOS⁃YOLOv5 algorithm improves the detection precision by 4%,the recall rate by 2.4%,the mean average precision by 3.1%and the loss function value by 5%in comparison with those of the traditional YOLOv5 model.The final detection accuracy of the improved algorithm is 95.38%.