Research On Monocular Image Depth Estimation

Monocular image depth estimation is research hotspot in the field of computer vision.However,restricted by a single camera,it is extremely difficult to calculate the accurate depth information in the image through spatial geometry.Monocular image depth estimation is to assign a relative depth relation to all the pixels in an image.Traditional methods usually use artificial features to estimate depth,which is prone to refactoring errors.In recent years,CNN have made remarkable achievements in the field of computer vision.Compared with artificial features,CNN features are acquired from large-scale data and can extract image features automatically and efficiently and represent rich semantic information of images.In this paper,CNN features are used to study the monocular image depth estimation.The specific contents are as follows:(1)Monocular image depth estimation algorithm based on CNN features extraction and weighted transfer learning.The depth estimation of monocular images can be obtained from similar images and their corresponding depth information.However,the performance of such an algorithm is limited by image matching ambiguity and uneven depth mapping.For this reason,CNN features are first introduced to obtain more accurate and effective similar image sets,and a monocular image depth estimation algorithm based on CNN features extraction is proposed.Firstly,CNN features are extracted to collect the neighboring image gallery of the input image.Secondly,pixel-wise dense spatial wrapping functions calculated between the input image and all candidate images are transferred to the candidate depth maps.The final depth image could be obtained by optimizing the transferred candidate depth maps.Then,aiming at the problem of uneven depth estimation of the same target in this algorithm,the authors have introduced the transferred weight SSW based on SIFT,and a monocular image depth estimation algorithm based on CNN features extraction and weighted transfer learning was further proposed.Experimental results show that CNN features extraction and weighted transfer learning improve the quality and accuracy of depth estimation.(2)Monocular image depth estimation algorithm based on multi-level CNN features fusion.The depth estimation of monocular image can be obtained by extracting the high-level features of training images through CNN network.However,the lack of local details leads to inaccurate depth estimation of targets in the scene.Therefore,a monocular image depth estimation algorithm based on multi-level CNN features fusion is proposed.The multi-level CNN features fusion network is used to extract not only the high-level features representing the spatial structure information of images,but also the low-level features representing the local detail information of images.The final depth image could be obtained by fusing these features of different levels and scales.Experimental results show that,the depth map estimated by this method can effectively maintain the local details of the scene.In this paper,we combined and applied the rich image feature expressions represented by CNN to the monocular image depth estimation,and the proposed method improved the performance of depth estimation.