论文部分内容阅读
针对从单目红外图像中恢复深度信息的问题,提出了一种基于深层卷积神经网络(DCNN)的深度估计方法。用劳斯掩膜和梯度检测器分别提取不同尺度下红外图像的纹理能量与纹理梯度,并将这两种纹理信息作为红外图像的第一种特征;提取图像中像元及其邻域的灰度值,以及统计其灰度直方图作为另外两种特征;分别用三种特征和深度信息标签训练DCNN,得到三种训练后的DCNN分别对单目红外图像进行深度估计。实验结果表明,相比较另外两种特征,用纹理信息训练的DCNN能够更有效地估计深度,并且优于现有的估计方法,尤其能较好地表现局部场景的深度变化。
Aimed at the problem of recovering depth information from monocular infrared images, a depth estimation method based on deep convolutional neural network (DCNN) is proposed. Using Raus mask and gradient detector, the texture energy and texture gradient of the infrared image at different scales were extracted respectively, and the two texture information were taken as the first feature of the infrared image. The gray value of the pixel and its neighborhood in the image As well as the histogram of the gray histogram as the other two features. The DCNN was trained by using three kinds of feature and depth information labels, respectively. Three kinds of trained DCNN were used to estimate the depth of the monocular infrared image respectively. Experimental results show that compared with the other two features, the DCNN trained by texture information can estimate the depth more effectively and is superior to the existing estimation methods, especially, it can better represent the depth variation of the local scene.