【摘要】 针对传统目标检测网络在低照度环境中的漏检及检测精度低的问题，提出一种基于SFFNet与YOLOv11的低照度目标检测网络SFF-YOLO。首先，提出一种空频域融合的图像增强网络SFFNet,将低照度图像与照度引导图合并后输入编码模块提取特征。然后，设计了双域融合网络DDFNet,通过空间域处理模块SPB提升图像亮度，并采用频域处理模块FPB修复局部细节，将融合后的空间-频域特征与最小通道约束图拼接后输入解码模块实现图像去噪。最后，设计了联合损失函数，对SFF-YOLO进行端到端联合训练，提升模型的泛化能力和目标检测性能。使用LOL-v2和ExDark数据集进行实验。实验结果表明，SFFNet在LOL-v2-Real和LOL-v2-Synthetic数据集上的PSNR分别为23.11和25.08,SSIM分别为0.851和0.936,相较于对比网络，展现出更出色的增强效果。SFF-YOLO在ExDark数据集上的检测精度达到80.4%,较YOLOv11提升了3.1%,检测速度为91.82帧/秒，实现了高精度的实时检测。更多还原

【Abstract】 To address the issues of missed detections and low accuracy in traditional object detection networks under low-light conditions, a low-light object detection network called SFF-YOLO is proposed. It′s based on SFFNet and YOLOv11. First, an image enhancement network called SFFNet is introduced. It employs spatial-frequency domain fusion, where low-light images are combined with illumination guidance maps and input into a coding module for feature extraction. Then, a dual-domain fusion network, DDFNet, is designed. Image brightness is enhanced through a spatial domain processing module(SPB), and local details are repaired using a frequency domain processing module(FPB). The fused spatial-frequency domain features are concatenated with a minimum channel constraint map and fed into a decoding module for image denosing. Finally, a joint loss function is designed. End-to-end joint training of SFF-YOLO is conducted through it, which improves the model′s generalization ability and object detection performance. Experiments are conducted on the LOL-v2 and ExDark datasets. The experimental results show that PSNR values of 23.11 and 25.08 as well as SSIM values of 0.851 and 0.936 are achieved by SFFNet on the LOL-v2-Real and LOL-v2-Synthetic datasets respectively, which demonstrates superior enhancement effects compared to competing networks. An accuracy of 80.4% is reached by SFF-YOLO on the ExDark dataset, which is an improvement of 3.1% over YOLOv11, with a detection speed of 91.82 frames per second, achieving high-precision real-time detection.更多还原