改进的YOLOv3算法对伪装目标检测

Camouflage Target Detection Based on an Improved YOLOv3 Algorithm

针对现有算法对伪装目标检测效率较低的问题,提出了一种基于改进YOLOv3网络的伪装目标检测算法。该算法在现有数据集下,对算法的先验框进行重聚类;依据Darknet53多次利用残差块的特点,将残差网络的级联方式由单级跳连改为多级跳连,改善误差在网络中的回传效果;在网络中添加了注意力模块,增加有用特征的检测权重。实验结果表明,与原始算法相比,改进后的算法模型具有较低的损失值且平均精度均值提高了约4.35%。

Aiming at the low efficiency of existing algorithms for camouflage target detection,a camouflage target detection algorithm based on the improved YOLOv3 network is proposed.Firstly,the algorithm re-clusters the priori boxes of the algorithm under the existing datasets;secondly,according to Darknet53 the characteristics of the residual block are utilized many times,the cascading mode of the residual network is changed from Single-stage jump to multi-stage jump connection to improve the return effect of errors in the network;finally,an attention module is added to the network to increase the detection weight of useful features.The experimental results show that compared with the original algorithm,the improved algorithm model has a lower loss value and the mean average precision is increased by about 4.35%.