Deep Learning Based Depth Recovery From Single Monocular Image

Three-dimensional reconstruction has always been one of the research hot topic in the field of computer vision,and the acquisition of depth information plays an extremely important role in three-dimensional reconstruction.The traditional method obtains depth information by using stereo vision and geometric relationship,which requires restricting environmental conditions and also has lots of limitations.Consequently,the more robust method of depth information recovery from single monocular image,which is closer to the actual scene than other traditional method has great research value.Because convolutional neural networks has a powerful talent of autonomous learning,this thesis proposes a deep learning based method,which recovers depth information from monocular image directly.Firstly,this thesis builds a novel convolutional neural network structure.The network splits the feature sampling into multiple scales,simultaneously performs rough sampling of global features and fine sampling of local features.Secondly,the network joins the improved DenseNet structure to improve feature gathering efficiency,which has great advantage of features propagation and features reuse through dense connection,finally realizes the multi-level comprehensive and efficient use of features.The experiments on NYU Depth V2 dataset illustrate the effectiveness of this method.The average relative error is 0.119,the root mean squared error is 0.547,and the average log10 error is 0.052.Besides,this thesis transplants the network to an embedded device and runs successfully,realizing a simple embedded artificial intelligence,but there is still much room for improvement in speed.In summary,this thesis applies the convolutional neural network to the algorithm of depth information recovery from monocular image.The proposed algorithm can directly acquire depth information form monocular images,which provides a new idea for three-dimensional reconstruction.In addition,the training method of supervised learning is adopted in the proposed model,further attempts will consider more unsupervised learning elements.