基于深度图去噪的RGBD显著性目标检测的研究

Research on RGBD Salient Object Detection Based on Depth Map Denoising

虽然基于深度图像(RGB-Depth map,RGBD)的显著性目标检测如今已取得了很大的发展,但是深度图含有大量的噪声,使得检测性能还有较大的提升空间。首先使用单目深度估计的方法预测出RGB图片的深度图,并与实际输入的深度图作MAE、E-measure、S-measure评分,可以选取质量较好的预测深度图;其次使用筛选的预测深度图训练一个深度图去噪网络,达到了对输入深度图进行去噪的目的;最后,将去噪的深度图、RGB图像做显著性目标检测与不去噪的深度图、RGB图像做显著性目标检测进行对比,前者更加优异,证实了深度图去噪对显著性目标检测性能提升的有效性。

Although significant target detection based on RGBD has made great progress, the depth map contains a lot of noise,which makes the detection performance still have a large room for improvement. Firstly, the monocular depth estimation method is used to predict the depth map of RGB image, and the MAE, E-measure and S-measure scores are made with the actual input depth map to select the prediction depth map with better quality;Secondly, the filtered predicted depth map is used to train a depth map denoising network to denoise the input depth map;Finally, the significant target detection of denoised depth map and RGB image is compared with that of non denoised depth map and RGB image. The former is more excellent, which proves the effectiveness of depth map denoising in improving the performance of significant target detection.