融合位置注意力的空洞卷积无人机影像多目标检测方法

A Multi-target Detection Method Based on Dilated Convolutional UAV Images Fused with Location Attention

针对以无人机影像为基础的多类别目标检测问题,提出一种融合位置注意力的卷积神经网络检测模型。在特征提取端使用级联空洞卷积核组实现特征提取,并通过位置卷积模块对提取后特征进一步筛选;为增强参与检测特征图内信息的复杂程度,通过多尺度特征融合端对上层输出特征图进行融合与上采样连接,最终输出4个尺度的特征图参与最终检测。测试结果显示,该模型在检测精度方面优于其余几组对比模型,同时在多个场景下表现出较好的泛化能力,在测试环境下能够实现对目标的实时检测。

Aiming at the problem of multi-category target detection based on UAV images,this paper proposed a convolutional neural network detection model fused with location attention.At the feature extraction end,the cascaded dilated convolutional kernel group was used to realize feature extraction,and the extracted features are further screened by the position convolution module;in order to enhance the complexity of the information in the feature map participating in the detection,the multi-scale feature fusion end is used to output features of the upper layer.The graphs are fused and connected by unsampling,and it finally outputs feature maps of four scales to participate in the final detection.The test results show that the model was superior to the other comparison models in terms of detection accuracy,at the same time shows good generalization ability in multiple scenarios,and can achieve real-time detection of targets in the test environment.