融合图像修复的遮挡目标检测算法

Occluded Target Detection Algorithm with Image Restoration

针对目标检测过程中存在遮挡导致的漏检和误检问题,提出一种融合图像修复模块的军事目标检测算法,通过深度学习实现军事目标的自动识别,再结合图像修复对遮挡目标实现图像增强。对于目标检测模块,在YOLOv4的基础上添加了卷积注意力机制,来增强对目标识别的敏感程度以及网络的特征提取能力;并且采用交叉迭代批量标准层,提高模型的训练效率。图像修复模块是基于生成对抗网络设计的一种双生成器模型,考虑到目标图像轮廓的完整性对图像的修复和目标的检测都有一定的影响,增加了一个边缘生成网络,图像修复模块旨在还原目标被遮挡的部位。模拟结果表明,融合了图像修复模块的目标检测算法对遮挡目标的检测精度达到了79.63%,较好地解决了遮挡情况下的漏检和误检问题。

In order to solve the problem of missed detection and false detection caused by occlusion in the process of target detection,a military target detection algorithm with image restoration module is proposed,which realizes the automatic recognition of military targets through deep learning,and then combines image restoration to achieve image enhancement of occluded targets.For the target detection module,a convolutional attention mechanism is added on the basis of YOLOv4 to enhance the sensitivity to target recognition and the feature extraction ability of the network.And the cross-iterative batch standard layer is adopted to improve the training efficiency of the model.The image restoration module is a double generator model designed based on the generation confrontation network.Considering that the integrity of the target image outline has a certain impact on image restoration and target detection,an edge generation network is added.The image restoration module aims to restore the part of the target that is occluded.The experimental results show that the target detection accuracy of the target detection algorithm with image restoration module reaches 79.63%,which effectively solves the problem of missed detection and false detection in the case of occlusion.