残差特征融合的小目标动态实时检测算法

针对图片中小目标携带信息少、尺度变化大等检测难点，本文以YOLOv5s为框架，提出一种特征融合的小目标动态实时检测模型(HCD-YOLOv5s)。针对模型下采样易造成小目标信息丢失、深层网络位置信息表达不足等问题，从浅层中新增检测小目标的检测头；本文针对特征融合造成的特征混淆等问题，设计一种特征融合方式CCAT,减少检测层位置信息与语义信息的丢失；针对检测任务与数据分布不同适应的激活函数不一致，设计DConv模块，分离回归任务与检测任务，实现模型的动态检测。本文在VisDrone数据集上对模型进行消融实验，3个模块相互促进。选取不同输入尺寸的图片对模型进行速度与精度测试。在YOLOv5s的基础上HCD-YOLOv5s的mAP50提高了10.2%,检测精度与参数量明显优于YOLOv5m, FPS达到90。最后在DOTA-v1.0上进行实验验证，mAP50、mAP分别提升了1.8%与2.0%,证明本文提出的HCD-YOLOv5s在小目标检测上有更佳的性能。

Small target dynamic real-time detection algorithm based on residual feature fusion

Aiming at the detection difficulties of small targets in pictures, such as less information and large scale changes, this paper proposes a feature fusion small target dynamic real-time detection model (HCD-YOLOv5s) based on YOLOv5s. In view of the problems that sampling under the model is easy to cause the loss of small target information and insufficient expression of deep network location information, a detection head for detecting small targets is added from the shallow layer; Aiming at the problem of feature confusion caused by feature fusion, this paper designs a feature fusion method CCAT to reduce the loss of location information and semantic information in the detection layer; In view of the inconsistency between the detection task and the activation function adapted to the different data distribution, the DConv module is designed to separate the regression task and the detection task, so as to realize the dynamic detection of the model. In this paper, the Ablation Experiment of the model is carried out on the visdrone data set, and the three modules promote each other. Select pictures with different input sizes to test the speed and accuracy of the model. On the basis of YOLOv5s, the mAP50 of HCD-YOLOv5s is increased by 10.2%, the detection accuracy and parameter quantity are significantly better than YOLOv5m, and the FPS reaches 90. Finally, the experimental verification is carried out on DOTA-v1.0, and the mAP50 and mAP are increased by 1.8% and 2.0% respectively, which proves that the HCD-YOLOv5s proposed in this paper has better performance in small target detection.