改进YOLOv5的遥感图像目标检测

Improved YOLOv5 Remote Sensing Image Target Detection

针对遥感图像目标识别过程中存在目标密集、目标遮挡、背景复杂等问题,提出一种改进的YOLOv5算法。首先,对锚框尺寸进行优化,使得到的每个锚框大小尺度更准确,有效提升了目标检测的准确度。其次,增加了卷积注意力机制,更关注感兴趣区域,抑制了无用信息,提高了算法的特征提取能力。最后,通过增加浅层特征图,提取学习目标特征,增加了网络对小目标的识别精度。数据集上验证表明,改进算法相较于YOLOv3、YOLOv4、Faster-RCNN、YOLOv5在识别精度上显著提升,在不同场景下也具有更好的鲁棒性。同时改进算法mAP达到97.0%,相比原始YOLOv5提升了2.2%。

Aiming at the problems of target density,target occlusion and complex background in the process of remote sensing image target recognition,an improved YOLOv5 algorithm is proposed.Firstly,the size of the anchor frame is optimized to make the size scale of each anchor frame more accurate and effectively improve the accuracy of target detection.Secondly,the convolutional attention mechanism is added to pay more attention to the region of interest,suppress useless information and improve the accuracy of target detection.Finally,by adding shallow feature graph to extract learning target features,the recognition accuracy of small targets is increased.Experimental results on data set show that the proposed algorithm has significantly improved recognition accuracy in comparison with YOLOv3,YOLOv4,Faster-RCNN and YOLOv5,and has better robustness in different scenarios.At the same time,the proposed algorithm mAP reaches 97.0%,which is 2.2%higher than original YOLOv5.