A homography matrix estimation method based on convolutional neural network

• Main Authors: DAI SHENGKUI, LIN CAIMING, WEI ZHIMIN, GAO JIANPING
• Format: Patent
• Language: Chinese; English
• Published: 11.12.2018
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention relates to a homography matrix estimation method based on a convolutional neural network. Firstly, a large number of data sets are generated. The data set is input into the convolutionalneural network to be trained. The network structure consists of 10 convolutional layers, 10 group normalization layers, 4 pooling layers, 2 full connection layers and 2 dropout layers. After two images with deformation transformation are input to the convolutional neural network, eight real numbers, namely homography matrix, are output at the last layer. The invention provides a convolutional neural network model estimation homography matrix method, which is an end-to-end estimation homography matrix mode, and provides a method for calculating the homography matrix of an image. 本发明涉及本发明种基于卷积神经网络的单应性矩阵估计方法，首先生成大量的数据集；将数据集输入构建的卷积神经网络中进行训练，网络结构共含有10个卷积层、10个群组归化层、4个池化层、2个全连接层和2个dropout层；将具有变形变换的两幅图像输入该卷积神经网络后，在最后层输出8个实数，即单应性矩阵；本发明提供的卷积神经网络模型估计单应性矩阵方法，是种端到端的估计单应性矩阵方式，为计算图像的单应性矩阵提供种方法。