基于改进Faster R-CNN的无人机小目标检测算法

Small target detection algorithm of drones based on improved Faster R-CNN

为提升两阶段目标检测算法Faster R-CNN对无人机小目标的检测识别效果,提出了一种改进的Faster R-CNN目标检测算法。对原始Faster R-CNN算法的特征提取网络进行改进,使用卷积层更少的ResNet-18作为骨干网络,减少算法的参数量;针对无人机目标特点,对Faster R-CNN算法的特征金字塔网络中的特征融合方法进行改进,增强了目标特征与背景特征的对比度;使用双线性插值的方法改善因感兴趣区域池化而造成的预测框偏离的问题。最后,在构建的低空无人机数据集上对改进的算法进行了实验验证,结果表明提出的改进Faster R-CNN目标检测算法检测速度达到了35.5帧/秒(FPS),较原始Faster R-CNN算法的15.8 FPS,速度提高了约一倍,且算法的平均精确度均值mAP提升了0.7%,有效提高了算法对无人机小目标的检测识别性能。

In order to improve the detection and recognition effect of the two-stage target detection algorithm Faster R-CNN on small target of drones,an improved Faster R-CNN target detection algorithm is proposed.In the improved algorithm,firstly,the feature extraction network of the original Faster R-CNN algorithm is improved,and ResNet-18 with fewer convolutional layers is used as the backbone network to reduce the number of parameters of the algorithm.Secondly,according to the characteristics of the drone target,the feature fusion method in the feature pyramid networks of Faster R-CNN is improved to enhance the contrast between target feature and background feature.Thirdly,the bilinear interpolation method is used to solve the problem of the deviation of the prediction frame caused by the pooling of regions of interest.Furthermore,the verification experiments are carried out on the constructed low-altitude drone data set.The results show that the improved Faster R-CNN target detection algorithm has a detection speed of 35.5 frames per second(FPS),which is about double the speed of the original Faster R-CNN algorithm(15.8 FPS),and the improved algorithm’s mean avergage precison(mAP)is increased by 0.7%,which effectively improves the detection and recognition performance of the algorithm for drone small targets.