Abstract: Due to the multiple down sampling operation, deep learning based monocular depth estimation always results in some problems such as lack of detailed information and blurred reconstructed edge contours.For these issues, a self-supervised monocular depth estimation method based on high resolution convolutional network is proposed. Firstly, to recover the detailed information, this work introduces the parallel connection in the encoder to keep the feature map in high-resolution. Secondly, in order to improve the learning ability of the encoder, the attention module is introduced to filter and refine the image features. Finally, aiming at the ambiguity of depth estimation, this work designs an effective loss function according to the consistency between non-adjacent frame images, and uses a mask to eliminate the interference of moving points and occluded points. We implement the depth estimation system based on the TensorFlow framework,and the training is carried out on the KITTI and Cityscapes datasets. Through comparative analysis with proposed depth estimation methods, the results show that this method can not only maintain the edge information of the predicted depth, but also improve the accuracy of the predicted depth, with an average relative error of 0.119.