基于改进YOLOv5的无人机视觉目标检测

UAV Vision Target Detection Based on Improved YOLOv5

针对无人机视觉下图像背景复杂、目标过小导致的检测效果过差的问题,提出了一种基于YOLOv5的改进算法。首先,提出滤波分离特征提取(FSFE)结构,将滤波分离后的图像与原图像并行输入神经网络进行特征提取,加强了网络对全局和细节的重要信息的提取,将输出特征图进行空间自适应融合,防止了融合时的语义信息割裂的问题,并且使得网络能够更加关注关键层的信息。其次,增加小目标检测层,并利用了SPD卷积模块加强特征学习来提高检测性能。最后,在C3模块中镶嵌CA特征增强模块,在特征提取时挖掘并保存重要的语义信息。基于VisDrone2019数据集的实验结果表明,改进算法的mAP@0.5和mAP@0.5∶0.95分别提升了8.3和6.1个百分点,精确率和召回率分别提升了5.1和4.5个百分点,提升了小目标检测精度,同时减少了漏检、误检概率,对实现无人机视觉小目标检测有重要意义。

An improved algorithm based on YOLOv5 is proposed to address the problem of poor detection due to complex image background and too small target under UAV vision.Firstly,the algorithm proposes a Filter Separation Feature Extraction(FSFE)structure,which inputs the filter separated image into the neural network in parallel with the original image,strengthens the network's extraction of important information both globally and in detail,the output feature map is spatially adaptively fused to prevent the problem of semantic information fragmentation during fusion,and enables the network to pay more attention to the information of key layers.Secondly,a small target detection layer is added,and the SPD convolution module is utilized to enhance feature learning to improve detection performance.Finally,the CA feature enhancement module is embedded in the C3 module to mine and preserve important semantic information during feature extraction.Experimental results based on the VisDrone 2019 dataset show that mAP@0.5 and mAP@0.5∶0.95 of the improved algorithm increases by 8.3 and 6.1 percentage points respectively,and the accuracy and recall increases by 5.1 and 4.5 percentage points respectively,improving the precision of small target detection and reducing the probability of missed and false detection,which is significant for realizing the UAV visual small target detection.