Depth Information Intelligent Based On Image Geometric Perception

As one of the most challenging issues in the field of computer vision,depth estimation can solve problems such as 3D modeling,scene understanding,depth aware image synthesis etc.Most of deep information detection algorithms in deep learning were based on the pixel-level or super-pixel-level features of the image to analyze the shape,texture,and color of the image.However,this method was easy to be affected by the external environment factors,and can not realize the rotation invariance of the image,and there was a serious problem of pixel feature redundancy.Line segments can be used as a basic linear feature to construct image semantic information,and it have rotation invariance.In this paper,the geometric features of the image are extracted by the line segment detection method,then we designed a line pre-processing method.On this basis,we proposed a line segment convolution neural network,using the seg-conv and seg-pool modules,which is used to implement the image geometric features Recognition algorithm and image depth line segment classification algorithm.By improving the line segment detection algorithm based on image segmentation,the effect of image vanishing point detection is also improved.The main contents of this paper are as follows:1.In this work,we processed the MNIST dataset using the LSD algorithm and the discrete curve evolution algorithm.Then we used sparse autoencoder to classify the new datasets.The results proved that line segment features can effectively restore the information of input image,and it is feasible to use line segment features to assist or complete image processing tasks independently.2.In this paper,we proposed a depth-segment classification algorithm based on a convolutional neural network.The NYU-Depth dataset was used by proposed algorithm to detect the segment-based features from original images.Afterwards,the line segments and labels of the character depth information obtained by the data preprocessing are combined with the depth data.The convolutional neural network was designed by cosidering the characteristics of the segments,and the classification of depth segments in monocular images was achieved.By conducting several multi-group comparison experiments on different hyper-parameters,it is proved that the convolution neural network can be used to realize the classification of depth segments,which contributed to solve the depth estimation problem by using the geometric features of the image.However,due to the large number of interference line segments in the image,the accuracy of model classification still has room for further improvement.3.This paper researched and improved the line segment detection algorithm based on image semantic segmentation.The mean-pooling and upsampling are used to instead of the dilated convolution,the activation function and loss function was replaced,used multi-scale of pooling for cascade training,which has improved time efficiency and detection results.The improved line segment detection algorithm is used to improvement the vanishing point detection network.The results proved that this method improves the accuracy and the efficiency of image vanishing point detection to a certain extent.