Statistics of Large, Medium, and Small Targets.

An MDCFVit-YOLO model based on the YOLOv8 algorithm is proposed to address issues in nighttime infrared object detection such as low visibility, high interference, and low precision in detecting small objects. The backbone network uses the lightweight Repvit model, improving detection performance and reducing model weight through transfer learning. The proposed MPA module integrates multi-scale contextual information, capturing complex dependencies between spatial and channel dimensions, thereby enhancing the representation capability of the neural network. The CSM module dynamically adjusts the weights of feature maps, enhancing the model of sensitivity to small targets. The dynamic automated detection head DAIH improves the accuracy of infrared target detection by dynamically adjusting regression feature maps. Additionally, three innovative loss functions—focalerDIoU, focalerGIOU and focalerShapeIoU are proposed to reduce losses during the training process. Experimental results show that the detection accuracy of 78% for small infrared nighttime targets, with a recall rate of 58.6%, an mAP value of 67%. and a parameter count of 20.9M for the MDCFVit-YOLO model. Compared to the baseline model YOLOv8, the mAP increased by 6.4%, with accuracy and recall rates improved by 4.5% and 5.7%, respectively. This research provides new ideas and methods for infrared target detection, enhancing the detection accuracy and real-time performance.

针对夜间红外目标检测中存在的能见度低、干扰强、小目标检测精度不足等问题，本文提出了一种基于YOLOv8算法的MDCFVit-YOLO模型。该模型的主干网络采用轻量级Repvit模型，通过迁移学习提升检测性能并降低模型权重。所提出的MPA模块融合多尺度上下文信息，捕捉空间维度与通道维度间的复杂依赖关系，进而提升神经网络的特征表征能力。CSM模块可动态调整特征图权重，增强模型对小目标的感知灵敏度。动态自动化检测头DAIH通过动态调整回归特征图，提升红外目标检测精度。此外，本文还提出了三种创新损失函数：focalerDIoU、focalerGIOU与focalerShapeIoU，以降低训练过程中的损失值。实验结果表明，MDCFVit-YOLO模型在夜间红外小目标检测任务中，检测准确率达78%、召回率为58.6%、平均精度均值（mean Average Precision, mAP）为67%，模型参数量为20.9M。相较于基线模型YOLOv8，该模型的mAP提升了6.4个百分点，准确率与召回率分别提升4.5%与5.7%。本研究为红外目标检测领域提供了全新的思路与方法，有效提升了检测精度与实时性。