应用残差稠密网络的无监督单幅图像深度估计

Unsupervised Depth Estimation from a Single Image Using Residual Dense NetWork

针对单幅图像深度估计中由于轮廓信息模糊造成的深度估计值不准确的问题,本文提出了一种应用残差稠密网络(Residual Dense Network)的单幅图像深度估计方法.该方法通过将残差稠密模块(Residual Dense Model)引入到具有跳跃连接(Skip connection)的编码器解码器结构,提出了一种新的神经网络模型.使用来自双目摄像机的一系列立体图像对,实现神经网络的无监督训练.通过将预测图像输入网络模型得到对应的视差图,再根据视差图与深度图之间的几何关系,得到图像的深度图.本文所提出的网络模型在KITTI驾驶数据集上进行训练,在测试集上得到了优于现存的大部分方法的误差值和准确率,以及更为清晰的物体边缘轮廓信息,从而验证了本文所提出方法在单幅图像深度估计中的有效性和优异性.

To overcome the inaccuracy of depth estimation value caused by fuzzy contour information in single image depth estimation,this paper proposes a single image depth estimation method using Residual Dense Network. This method proposes a new neuralnetwork model by introducing the Residual Dense Model into the encoder decoder structure with Skip connection. Unsupervised training of neural networks is achieved by using a series of stereo image pairs from binocular cameras. The disparity map is obtained by inputting the predicted image into the network model,and according to the geometric relationship between the disparity map and the depth map,the depth map of the image is obtained. The network model proposed in this paper is trained on the KITTI driving dataset,and the results show that the method obtains lower error value,higher accuracy and clearer edge contour information of the object than most of the existing methods on the test set. It proves the effectiveness and superiority of this paper in single image depth estimation.