基于改进YOLOv11n的短波信号检测算法

AFM-YOLOv11: Shortwave Signal Detection Algorithm Based on Improved YOLOv11n

摘要:

为提升复杂电磁环境下短波信号的检测精度，解决信号特征微弱、背景噪声干扰严重、信号尺度差异大导致的检测精度不足问题，对基于YOLOv11n的短波信号检测算法进行了研究。提出了一种改进的短波信号检测算法AFM-YOLO。算法在主干网络中将AssemFormer模块与C3K2深度融合，构建C3K2_AssemFormer模块，结合卷积的局部特征提取能力与Transformer的全局上下文建模优势，以增强对微弱信号的特征保持能力；引入了基于频域的自注意力求解器FSAS模块，利用频域计算的高效性降低注意力机制复杂度，提升模型在噪声环境下的特征区分能力；在颈部网络设计了多分支辅助特征金字塔MAFPN结构，通过双向融合机制强化浅层细节信息与深层语义信息的交互，以提升多尺度信号的检测能力。在包含ANDVT、2GALE、LINK11等10类短波信号的自建数据集上进行了实验验证。结果表明，AFM-YOLO的mAP50与mAP50-95分别达到93.5%和88.3%，较基准模型YOLOv11n提升了3.7和10.6个百分点，查准率提升3.5个百分点至94.5%，召回率提升5.9个百分点至91.2%。该算法能有效提升复杂电磁环境下短波信号的检测精度。

Abstract:

To improve the detection accuracy of shortwave signals in complex electromagnetic environments and address the issues of weak signal features, severe background noise interference, and significant signal scale variations that lead to insufficient detection precision, a shortwave signal detection algorithm based on YOLOv11n was investigated. An improved shortwave signal detection algorithm, named AFM-YOLO, was proposed. The algorithm deeply integrates the AssemFormer module with C3K2 in the backbone network to construct the C3K2_AssemFormer module, combining the local feature extraction capability of convolutions with the global context modeling advantage of Transformers to enhance feature preservation for weak signals. A frequency-domain-based self-attention solver (FSAS) module was introduced, leveraging the efficiency of frequency-domain computation to reduce the complexity of the attention mechanism and improve the model's feature discrimination capability in noisy environments. A multi-branch auxiliary feature pyramid network (MAFPN) structure was designed in the neck network, employing a bidirectional fusion mechanism to strengthen the interaction between shallow-layer detail information and deep-layer semantic information, thereby improving the detection capability for multi-scale signals. Experimental validation was conducted on a self-built dataset containing ten types of shortwave signals, including ANDVT, 2GALE, and LINK11. The results indicate that AFM-YOLO achieves mAP50 and mAP50-95 scores of 93.5% and 88.3%, respectively, representing improvements of 3.7 and 10.6 percentage points over the baseline YOLOv11n model. The precision rate increased by 3.5 percentage points to 94.5%, and the recall rate increased by 5.9 percentage points to 91.2%. The proposed algorithm can effectively enhance the detection accuracy of shortwave signals in complex electromagnetic environments.