基于卷积神经网络的水下地形测量误差校正方法  

Error correction in underwater topographic survey based on convolutional neural network

作　　者：李世均[1] LI Shijun(Chongqing Survey Institute,Chongqing 401121,China)

机构地区：[1]重庆市勘测院,重庆401121

出　　处：《测绘技术装备》2024年第2期28-33,共6页Geomatics Technology and Equipment

摘　　要：水下地形测量精度对于水下地貌的判断及水下航行安全具有重要意义,测量误差会直接影响对水下地形的判断。因此,本文提出基于卷积神经网络的水下地形测量误差校正方法,对水下地形测量数据误差进行自动识别和校正,从而改善水下地形测量结果。该方法首先依据水下地形测量原理,统计水下地形测量的特征参数,形成特征向量;其次,将其作为基于改进卷积神经网络的输入,通过不断学习和训练,输出变换参数结果,生成新的水下地形测量图像;最后,引入多尺度注意力机制,细化测量图像空间,并对测量图像与标签图像之间的相似度进行计算,依据最大化图像相似度计算结果,校正水下地形测量图像生成过程中的参数。测试结果表明,该方法测量误差均在1.7%以下,能够有效校正测量结果中存在的畸变和误差,精准测量水下地形情况,测量结果与实际结果吻合程度较高。The accuracy of underwater topographic survey is of great significance to the judgment of underwater topography and navigation safety.The survey error directly affects the judgments of underwater topography.Therefore,the method for error correction in underwater topographic survey is proposed in this paper based on convolutional neural network,it can automatically identify and correct the errors in underwater topographic survey data,and improve the accuracy of underwater topographic survey.Firstly,the characteristic parameters of underwater topographic survey are counted according to the principles of underwater topographic survey,and the feature vectors established.Then it is used as the input based on the improved convolutional neural network to generate a new underwater topographic survey image through continuous learning and training.Finally,the multi-scale attention mechanism is introduced to refine the measurement image space,and the similarity between the survey image and the label image is calculated.The parameters in the process of underwater topographic survey image generation are corrected according to the maximum image similarity calculation results.The results show that the error is less than 1.7% when it is corrected,it can effectively correct the distortion and error,accurately survey the underwater terrain,and the surveyed results are highly consistent with the real conditions.

关 键 词：卷积神经网络 水下地形 测量误差校正 图像空间 细化测量 特征参数 

分 类 号：TP73[自动化与计算机技术—检测技术与自动化装置]