约束性神经网络及其在目标方向估计中的应用

Constrained Neural Network and Its Application in Target Direction Estimation

摘要: 目前，在计算机视觉领域，主流的卷积神经网络算法专注于目标的识别和定位，且大多数采用轴对齐包围盒定位目标，而为了对图像做更深刻的语义理解，更加精准地定位各类目标，需要获取目标的方向信息。因此，本文提出一种针对图像目标方向估计的方法，采用卷积神经网络对描述目标方向的两个方向角分量进行回归，规避一些现有方向估计方法直接对方向角回归而产生的缺点。由于方向分量之间存在平方和为1的函数约束，本文提出约束性神经网络的概念，进一步提出利用约束性神经网络解决这类带有输出约束问题的一般性方法，即在Loss层引入约束误差，参与反向传播，并将其具体运用于目标方向估计中。经实验，本文采用的基于约束性神经网络的目标方向估计方法，能够在保证原输出损失的下降速度和幅度的前提下，降低约束误差，提高估计精度。

Abstract: At present, the mainstream convolutional neural network (CNN) approaches focus on the recogni-tion and positioning of targets in the field of computer vision. Most locate the position of targets with axis-aligned bounding boxes. In order to make a deeper understanding of the images and obtain more accurate position of various targets, the direction information is needed. Therefore, a new method is proposed to estimate the target direction, which applies CNN to regress the two di-rectional components of the target direction angle that describes the direction, instead of directly regressing the direction angle like some existing approaches which shows some shortcomings. Considering the function constraint of the square sum of 1 between the directional components, this paper proposes a general method to solve this kind of problems with output constraints by using constrained neural network, which introduces constraint errors into the Loss layer and the back propagation process, and applies it in the target direction estimation. Experiments show that, the target direction estimation method maintains the descent rate and range of the original output loss and moreover, reduce the constraint error and improve the prediction accuracy. The significance lies in proposing a general method to solve a kind of problems with output constraints, which reduces constraint errors, and shows general applicability.