改进YOLOv4的铁路沿线遥感影像地物检测方法

近年来，高分遥感影像技术的快速发展为铁路沿线地物检测提供了一种重要技术手段。基于回归的一阶段目标检测方法YOLOv4具有检测精度高、速度快等优点，但用于遥感影像检测时仍然存在部分细节特征信息丢失导致的小目标漏检，以及进行大面积地物检测时效率低的问题。为此，提出改进YOLOv4网络模型对遥感影像铁路沿线地物进行检测。首先，设计由卷积、批量归一化和Mish激活函数组成的CBM(convolution batch normalization mish)模块，并采用DCBM(double CBM)模块作为密集连接网络(DenseNet)的传输层用于YOLOv4网络特征提取以实现地物特征传递和信息重用，增强小目标地物的检测能力，降低漏检率；然后针对YOLOv4在大面积检测时效率不高和模型参数空间较大的缺陷，将压缩激励SE(squeeze excitation)通道注意机制用于骨干网中跨阶段局部单元(cross stage partial, CSP)的每个残差单元之后，减少SE注意模块的重复调用次数，使其能够在提高网络性能的同时降低模型参数量从而提高检测效率；最后，针对长条形状的铁路目标提取困难问题，在网络结果输出之前引入改进的通道空间注意力机制ICBAM(improved convolutional block attention module) 保留原始特征信息，解决铁路目标特征提取能力差的问题，提高铁路中大尺度目标的检测效率。为验证所提方法的有效性，选取2 048张分辨率为1 920×1 080的某段铁路沿线遥感影像地物样本数据，将其中的铁路、房屋、楼宇建筑、农田和水池作为检测目标进行实验，并与当前流行的目标检测方法进行对比。结果表明，改进方法不仅增强了对小目标地物的检测能力，提高了地物检测精度和速度，而且提高了大尺度目标的检测效率。与YOLOv4算法相比，mAP提高了2.11%，准确率提高了2.93%，召回率提高了3.79%，模型大小减少了8.53%。所提方法为当前应用高速铁路沿线遥感影像地物快速精准检测提供了有效方法。

Improved YOLOv4 Remote Sensing Image Detection Method of Ground Objects Along Railway

In recent years, the rapid development of high-resolution remote sensing technology has provided an effective technical means for detecting ground objects along the railway. The regression-based one-stage target detection method YOLOv4 has the advantages of high detection accuracy and fast speed. However, when it is used for remote sensing image detection, small targets are still missed due to the loss of some detailed feature information, and large-area ground object detection. Due to the problem of low efficiency, this paper improves the YOLOv4 network model to detect ground objects along the railway in remote sensing images. This paper improves the YOLOv4 network model to detect the ground features in remote sensing images along the railway. First, the CBM (Convolution Batch Normalization Mish) module is designed with composing of convolution, batch normalization, and Mish activation, and the DCBM (Double CBM) module is used for the transmission layer of the densely connected network (DenseNet) for the YOLOv4 network feature extraction. It can achieve feature transfer and information reuse and enhance small target feature detection capabilities. Then, to address the defects of YOLOv4 in the inefficiency of large area detection and the large space of model parameters, the SE (Squeeze Excitation) channel attention mechanism is used after each residual cell of Cross Stage Partial (CSP) in the backbone network to reduce the number of repeated calls of the SE attention module. Hence the performance of the network is improved while reducing the number of model parameters and improving detection efficiency. Finally, for the problem of difficult extraction of railroad targets in images, an improved channel space attention mechanism ICBAM (Improved Convolutional Block Attention Module) is introduced before the network result output to retain the original feature information. It can solve the problem of poor feature extraction ability of railroad targets, and improve the detection efficiency of large-scale targets. To verify the effectiveness of the proposed method, 1 676 remote sensing image samples are selected along a particular section of the railway with a resolution of 1 920×1 080. Railways, houses, buildings, farmland, and ponds in the data set are selected as targets for inspection, and some current popular target detection methods are compared. The experimental results show that the improved method enhances the detection ability of small targets, improves the accuracy and speed of detection, and improves the detection efficiency of large-scale targets. Compared with the YOLOv4 algorithm, the improved method mAP has increased by 2.11%, accuracy increased by 2.93%, the recall rate has increased by 3.79%, and the model size is reduced by 8.53%. The proposed method also provides an effective method for rapidly and accurately detecting ground objects in remote sensing images along the high-speed railway.