一种基于CNN的散射变换的旋转方法

A Rotation Method of Scattering Transformation Based on CNN

深卷积神经网络的精度与输入的数据有关,当输入的数据为非正常角度数据时,相关网络的处理精度较低。因此,为了提高处理精确度,提出了一种对输入旋转不变的卷积神经网络。该网络主要基于散射变换中存在的旋转变换特性和一系列的三维卷积。这种结构能够在无角度标注数据的情况下预测角度。此外,预测器可以将输入的随机旋转连续映射到预测的圆形空间。在实验中,利用垂直角度和随机旋转角度的样本对训练结果进行了验证。结果表明该网络具有较好的精确度。

The accuracy of the deep convolutional neural network is related to the input data. When the input data are with abnormal angles, the processing accuracy of the related network is low. In order to improve the processing accuracy, a convolutional neural network that is invariant to the input rotation is proposed. The network is mainly based on the rotation transformation characteristics existing in the scattering transformation and a series of three-dimensional convolutions. This structure can predict the angle without angle annotation data. In addition, the predictor can continuously map the input random rotation to the predicted circular space. In the experiment, the training results are verified using samples of vertical angles and random rotation angles. The results show that the network has good accuracy.