Reasearch On Depth Estimation Method Based On Monocular Gastrointestinal Image

Gastroenteritis is a common disease in our life.At present,the diagnosis and treatment of a common gastrointestinal disease are by introducing gastroscope for observation.Depth estimation is an important means of providing distance information in the endoscope navigation system.However,the accuracy of depth estimation is greatly challenged by tissue diversity and vascular texture.To provide distance information better,this thesis explores depth estimation methods from two aspects: supervised learning and self-supervised learning.The supervised learning method used in this project requires the depth map of the image to be provided for training,and the convolutional neural network is used to realize depth estimation.Two ways of supervised training are proposed.One way is to realize supervised learning by using monocular gastrointestinal images,and the ber Hu loss function proposed is used as the supervised object.The second way is to use binocular gastrointestinal images to achieve supervised learning and proposes the consistency loss function of the left and right perspectives.Although binocular gastrointestinal images were used in the training,monocular images were still used in the testing stage and practical application.The principle of the left-right perspective consistency loss function is to transform the binocular depth map to the same coordinate system to minimize the difference between the two.The above two supervised methods both use multiscale loss calculation and a new decoding structure.Besides,we introduce the attention mechanism into the depth estimation task to improve the accuracy of depth estimation.Specifically,we redesign the channel attention mechanism and spatial attention mechanism and embed them reasonably into the convolutional neural network.In this thesis,a depth estimation method of self-supervised learning is proposed,mainly in order to train and learn to predict depth maps when depth map labels are insufficient or cannot be provided.Specifically,we use binocular vision reconstruction to construct the loss function.If we can reconstruct an image from one perspective to another,then we have the depth information.The prediction object in this part is converted from a depth map to a disparity map.In practical application,a disparity map can be converted into a depth map by using camera calibration parameters.The structure of the convolutional neural network used in self-supervised learning is the same as that used in supervised learning,and attention mechanism is added to improve the prediction accuracy.In the model testing stage,the same distributed test set and different distributed test sets are used to explore the performance of the model.In the test results of the same distribution test set,the average absolute error is the optimal 6.49 mm,which is better than the different distribution test set.In the test results of disparity error using self-supervised learning,the results of the same distribution test set are better than those of the different distribution test set.In supervised learning and self-supervised learning,the attention mechanism can improve the performance of the model.