结合多通道深度学习和随机森林的地表分类

Surface Classification Combining Multi-Channel Deep Learning With Random Forest

地表分类技术对地面无人驾驶车辆的感知能力有着重要影响。而针对传统卷积神经网络CNN(Convolutional Neural Networks)地表分类效果不佳的问题,本文提出一种结合多通道深度学习和随机森林的地表分类算法。算法先通过图像计算得到人工设计的特征LBP;再采用多通道融合技术,将原彩色图像的RGB三通道和LBP通道加以融合形成融合图像;然后构建并预训练卷积神经网络,以此提取融合图像的关键特征信息;最后用随机森林分类器代替卷积神经网络输出层完成地表分类。实验结果表明,本文算法识别正确率达到98.56%,相比于传统卷积神经网络能取得更好的分类结果,具有一定的鲁棒能力。

Surface classification technology has an important effect on perception ability of ground driverless vehicles.To solve the problem of ineffective land surface classification of traditional Convolutional Neural Networks(CNN),the paper proposes a land surface classification algorithm based on multi-channel deep learning and random forest.The algorithm firstly obtains artificially designed feature LBP by image calculation,then fuses RGB three-channel and LBP channel of original color image to form the fusion image with multi-channel fusion technology,constructs and pretrains convolutional neural network to extract key feature information of the fusion image,finally replaces output layer of convolutional neural network with random forest classifier to complete surface classification.Experimental results show recognition accuracy of the algorithm reaches 98.56%.Compared with traditional convolution neural network,the algorithm can achieve better classification results and has some certain robustness.