Unsupervised Monocular Depth Estimation Based On Deep Learning

With the continuous development of Convolutional Neural Networks,the excellent performance of deep learning algorithms in the field of computer vision has attracted many researchers.For depth estimation of 2D images,deep learning algorithms have gradually replaced the traditional methods which use expensive equipment to model the surface of objects and have become the most widely studied and applied method now.According to the training method,depth estimation based on deep learning methods can be categorized as supervised learning,unsupervised learning and semisupervised learning depth estimation.Depth estimation based on supervised learning is the current popular algorithm.However,it is expensive to obtain groundtruth labels.As an unsupervised deep learning model,Generative Adversarial Network(GAN)can learn to extract features by itself,and continuously optimize the model through the two-player game of the generator and the discriminator,without groundtruth as supervision.This unsupervised framework with high efficiency and low cost is widely used in the field of computer vision.However,the accuracy of GAN which performs on outdoor depth estimation is generally low,making it difficult to accurately determine the boundaries of objects.To solve this problem,this thesis proposes an improved Cycle Genarative Adversarial Network to improve the accuracy of monocular depth estimation in outdoor scenes.The main topics and innovation are as follows:(1)Aiming at the problem of low depth estimation accuracy of traditional GAN models,this thesis proposes ASPP-CycleGAN that contains a spatial pyramid structure.The framework,which consists of two generators and discriminators,uses an unsupervised learning method to transform the depth estimation problem into an image reconstruction problem.The structure of the generator is Encoder-ASPP feature extractor-Decoder.The ASPP(Atrous Spatial Pyramid Pooling)acts as a "bridge" between the encoder and the decoder,extracts information on multiple scales,effectively retains detailed information,and reduces the computational cost.Experimental results show that compared with the existing deep learning algorithms,the ASPP-Cycle GAN model has significantly improved performance on monocular depth estimation.(2)In view of the large number of GAN parameters,low training efficiency,and the “Dead ReLU Problem” during the training process,this thesis uses the ELU activation function in the activation layer of the discriminator network to replace the traditional ReLU activation function,effectively avoiding the parameters fail to update in the negative area.In addition,in the inactive state,ELU activation function is more robust to noise.Since the ELU activation function has normalization characteristics to the data,this thesis removes the instance normalization layer in the discriminator network,reduces the model parameters and effectively improves the calculation efficiency.Experimental results show that ASPP-Cycle GAN,which introduces the ELU activation function,solves the "Dead ReLU Problem",improves the performance of depth estimation and reduces the time for parameter update once by 0.005s.(3)Aiming at the problem that the accuracy of monocular depth estimation is not good enough and it is difficult to retain detailed information,this thesis proposes to use Post-Processing to optimize the disparity map generated by the previous model.Experimental results show that the ASPPCycleGAN based on Post-Processing operation can significantly improve the effect of depth estimation and effectively retain the detailed information in the input image.