卷积神经网络在航测图像自动识别中的应用探讨

Discussion on the Application of Convolution Neural Network in Automatic Recognition of Aerial Survey Image

针对无人机航测影像的目标识别问题,结合目前已有相关开发语言及模型,探讨在航测内业采集过程中加入人工智能识别技术实现地物自动识别和绘制的可行性。首先,分析近年来计算机图像识别方面的人工智能模型,结合航空影像固有特性,通过研究识别后与已有绘图软件交互。其次,设计了一组基于经典卷积神经网络的航测影像自动识别实验。结果表明,VGG16模型能够有效提升高分辨率和复杂背景的航拍图像的识别准确率,在较小目标(如路灯等)的识别准确率较低。以此给出输入图像精细化预处理、原数据集数据增强与多次迭代、构建具有双重损失函数的糅合模型3个方面的改进措施,为后续进一步的研究确定了方向。

Aiming at the object recognition problem of UAV aerial survey image,combined with the existing development language and model,the feasibility of adding artificial intelligence recognition technology to the process of aerial survey internal business acquisition to realize automatic recognition and rendering of ground objects was discussed.Firstly,the artificial intelligence model of computer image recognition in recent years is analyzed,and combined with the inherent characteristics of aerial image,through the study of recognition and interaction with the existing graphics software.Secondly,a group of automatic recognition experiments of aerial survey images based on classical convolutional neural networks arc designed.The results show that the VGG16 model can effectively improve the recognition accuracy of aerial images with high resolution and complex background,while the recognition accuracy of small targets such as street lights is low.In this paper,three improvement measures of input image refinement preprocessing,original data set data enhancement and multiple iterations,and the mash-up model with double loss function arc given.The direction is determined for the follow-up further research.