改进YOLOv3的红外弱小目标检测

Improved YOLOv3 in infrared dim small target detection

为解决在红外场景下小目标携带的特征信息较少,导致检测结果精度较低且容易出现漏检等问题,建立一种红外弱小目标检测模型。使用改进的K-means聚类算法对YOLOv3的anchor进行重新聚类,聚类中心点的迭代以交并比代替原来的欧氏距离。通过改进的空间金字塔池化模块将浅层空间特征与深层语义特征相融合,丰富红外弱小目标的特征信息。将EIoU引入到YOLOv3中,使目标框和锚框的宽度和高度的差异最小化。实验结果表明,该模型在SAITD数据集上的查准率达到了94.83%,平均查准率达到了89.26%,检测精度优于传统红外目标检测网络及部分深度目标检测网络。

To solve the problem that small targets carry less characteristic information in infrared scene,which leads to low accuracy of detection results and is easy to miss alarms,an infrared dim small target detection model was established.The improved K-means clustering algorithm was used to recluster the anchor of YOLOv3,and for the iteration of the cluster center points,the original Euclidean distance was replaced by the intersection ratio.Through the improved space pyramid pool module,the shallow spatial features and deep semantic features were integrated to enrich the feature information of infrared dim target.EIoU was introduced into YOLOv3 to minimize the difference in width and height between the target frame and the anchor frame.Experimental results show that the accuracy of the model on the SAITD dataset reaches 94.83%,and the average accuracy reaches 89.26%,which is better than that of the traditional infrared target detection network and some deep target detection networks.