Abstract:

In order to solve the problem that the traditional feature extraction method is insufficient in aircraft detection accuracy and real-time in remote sensing images, two improvements are proposed based on the YOLOv3-tiny network in terms of accuracy improvement. The improvement point one： the way to extract image feature points from the network improved to use packet convolution, that is, an image is divided into three channels for convolution operations, combined with channel feature transformation to enhance the semantic association between channels. Improvement point two： add a scale detection to the deep feature of the network, and perform the fusion prediction between the upsampling and the shallow feature map. Introducing deep separable convolution in the aspect of speed increase insteads of traditional convolution to reduce the amount of parameter calculation and achieve model weight reduction. According to the improved network, a modified convolutional neural network DS-YOLO（Depthwise Separable YOLO） with 33 convolutional layers is proposed. The improved network is trained on the self-made remote sensing aircraft image to select the optimal weight. It is used to test and analyze low-quality test sets such as small targets, high exposure, and background interference. Experiments show that the improved algorithm improves the accuracy of the test set by 14.1%, the recall rate by 16.8%, and the detection of low-quality remote sensing aircraft images.

Key words: depth separable convolution, packet convolution, DS-YOLO model, channel feature transformation, multi-scale prediction

摘要：

为了解决传统特征提取方法在遥感图像中飞机检测准确率和实时性不足的问题，基于YOLOv3-tiny网络在准确率提升方面提出两点改进。改进点一：将网络提取图像特征点的方式改进为分组卷积，即将一幅图像分成三个通道进行卷积操作，配合通道特征变换以加强各通道之间的语义关联；改进点二：将网络深层特征增加一个尺度检测，并进行上采样与浅层特征图进行融合预测。在速度提升方面引入深度可分离卷积代替传统卷积以降低参数计算量，达到模型轻量化。根据改进后的网络提出一种包含33个卷积层的改进型卷积神经网络DS-YOLO，对改进前后网络分别在自制遥感飞机图像上进行训练，选出最优的权重，用来对目标小、曝光度高、背景干扰等低质量测试集进行测试分析。实验结果表明，改进后的算法在测试集上精准度提升了14.1%，召回率提升了16.8%，检测低质量遥感飞机图像效果更佳。

关键词: 深度可分离卷积, 分组卷积, DS-YOLO模型, 通道特征变换, 多尺度预测