Monocular Image Depth Estimation Based On Deep Learning Model

As an important part of 3D scene analysis,monocular image depth estimation is also a hot research topic in the field of computer vision in recent years.Its goal is to estimate the depth of field of each pixel based on the monocular visible light map,that is,the distance of the scene from the camera.This is a challenging task.This thesis divides the monocular image depth estimation task into two steps,rough estimation and precision estimation.In the rough estimation stage,this thesis improves the FC-DenseNet deep convolutional neural network structure with dense block as a unit to simulate the fuzzy mapping between monocular image and depth map.In order to further improve the quality of the FC-DenseNet output depth map,superpixel algorithm SLIC is used to segment the monocular image in the second step,and a network structure that depends on the similarity of adjacent superpixel blocks CNN model called NSW-CNN and CRFasRNN module is proposed.The process firstly extracts three features of the local binary pattern LBP features,image color difference features,image color histogram distribution difference features in the super pixel block of the scene RGB map,and then normalizes the feature maps under the three characteristics.In this way,the depth map of the rough estimation output is linearly filtered,and then this filtering result is input a CNN network for further image enhancement as a joint filter.The CNN self-adaptive updating weight value calculates an optimal feature map that combines similarity factors of three neighboring super pixel blocks to enhance the three-dimensional scene information in the image,thereby improving the global accuracy of the depth map.Finally,this preliminary optimized depth map is input into the CRFasRNN module.In the joint end-to-end training process of the NSW-CNN,both the advantage of the conditional random field CRF and CNN are combined so that pixel-level optimization of the depth map is completed by the powerful modeling capability of the probability map model.