针对现有算法特征学习能力欠佳、检测精度不高、计算量大等问题,提出一种基于YOLOV5s改进的多尺度目标检测算法AEM-YOLOV5(AFPN-EMA-MPDIoU-YOLOV5)。首先,在颈部网络引入AFPN渐进特征金字塔网络,以渐进的方式融合图像底层详细信息和顶层高级语义特征,增强了网络特征融合效果;其次,在每个检测分支前增添EMA注意力机制模块,跨空间聚合像素级特征,提高了复杂场景下对多尺度目标的关注程度;最后,使用MPDIoU替代YOLOV5原有C_IoU边界框损失函数,解决了预测框宽高比相同但绝对值不同时C_IoU退化的问题,使回归结果更为准确。实验结果表明,改进后算法在RSOD数据集上P_mAP50达到94.5%,FPS达到42 frame/s,模型大小为14.8 MB。与现有算法相比,改进后算法性能显著提升,可满足军事目标检测的实时性要求、模型轻便。

Accurately and quickly detecting military targets in complex scenarios has important military value in perceiving battlefield situations, conducting reconnaissance and early warning analysis, and providing precise missile guidance. A multi-scale object detection algorithm AEM-YOLOV5 (AFPN-EMA-MPDIoU-YOLOV5) based on improvements to YOLOV5s is proposed to address the issues of poor feature learning ability, low detection accuracy, and high computational complexity in existing algorithms. Firstly, the AFPN asymptotic feature pyramid network is introduced into the neck network to gradually fuse the detailed information at the bottom of the image and the high-level semantic features at the top, enhancing the network feature fusion effect. Secondly, an EMA attention mechanism module is added before each detection branch to aggregate pixel level features across spaces, improving the level of attention to multi-scale targets in complex scenes. Finally, MPDIoU is used to replace the original C_IoU bounding box loss function in YOLOV5, solving the problem of C_IoU degradation when the predicted box aspect ratio is the same but the absolute value is different, making the regression results more accurate. The experimental results show that the improved algorithm performs well on the RSOD dataset, P_mAP50 reaches 94.5%, FPS reaches 42 frame/s, and model size is 14.8 MB. Compared with existing algorithms, the improved algorithm significantly improves its performance, meets the real-time requirements of military target detection, and the model is lightweight.