融合卷积神经网络与多层感知器的鞍部识别方法

针对传统鞍部识别方法中特征选择困难及未考虑鞍部与其它地形要素的共生关系等问题,利用深度卷积神经网络的特征自学习性能,提出了一种卷积神经网络与多层感知器相结合的混合模型实现DEM数据中的鞍部要素识别.首先设计改进的卷积神经网络模型自动提取鞍部的深度特征,经过Softmax分类器得到候选鞍部点,再运用多层感知器对候选鞍部点的位置进行精细回归,标识出最终的鞍部要素坐标.通过自建的鞍部样本集SADDLE-100训练网络模型,并在三种不同的山地样区进行实验,实验结果表明该方法比其它鞍部识别方法的漏提率减少约50%,正确识别率提高6.7%,在一定程度上避免了人工选择特征造成的鞍部语义信息缺失现象,为DEM中的点状要素识别提供了新的技术途径.

Saddle Recognition Method Based on Convolutional Neural Network and Multi-layer Perceptron

In view of the difficulty of feature selection in traditional saddle recognition method and the problem of not considering the symbiosis between saddle and other terrain elements,a hybrid model of convolution neural network and multi-layer perceptron is proposed to realize saddle element recognition in DEM data by using the feature self-learning performance of deep convolution neural network.Firstly,the improved convolutional neural network model is designed to extract the depth features of saddle automatically,and then the candidate saddle points are obtained by Softmax classifier.Then,the multi-layer perceptron is used to make precise regression for the position of the candidate saddle points,and the final saddle element coordinates are identified.Through the saddle sample set SADDLE-100 training network model and experiments in three different mountain sample areas,the experimental results show that this method can reduce the missing rate of other saddle recognition methods by about 50%and improve the correct recognition rate by 6.7%.To a certain extent,it can avoid the lack of saddle semantic information caused by artificial feature selection,and provide a new point element recognition in DEM Technical approach.