内窥镜图像深度估计是微创手术中的重要技术难题。为提高内窥镜图像深度估计的准确性，提出一种基于改进Monodepth2的内窥镜图像深度估计方法。在深度估计网络中，编码器使用ResNet34模块，并引入SAB（sparse attentive backtracking）注意力机制、改进的FPN（feature pyramid network）模块以及特征增强模块，以使所提网络更好地理解全局信息、灵活有效地处理多尺度特征，并进一步增强其稳定性和可靠性。解码器通过上采样获取图像的深度信息和位姿信息。采用光度重投影误差、结构相似性和边缘感知平滑误差作为损失函数，以进一步提高所提方法的准确性。评估采用Hamlyn公共数据集，实验结果表明：所提方法可更加准确地估计内窥镜图像的深度信息，进一步验证了所提方法的有效性和准确性。

Endoscopic image depth estimation is an important technical challenge in minimally invasive surgery. In order to improve the accuracy of depth estimation of endoscopic image, an endoscopic image depth estimation method based on improved Monodepth2 was proposed. In the depth estimation network, the encoder used the ResNet34 module and introduced the sparse attentive backtracking (SAB) attention mechanism, the improved feature pyramid network (FPN) module and feature enhancement module, so that the proposed network could better understand the global information, deal with multi-scale features flexibly and effectively, and further enhanced its stability and reliability. The decoder obtained the depth information and pose information of the image through up-sampling. To further improve the accuracy of the proposed method, photometric reprojection error, structural similarity and edge perception smoothing error were used as loss functions. The evaluation used Hamlyn public dataset, and the experimental results show that the proposed method can estimate the depth information of endoscope image more accurately, which further verifies the effectiveness and accuracy of the proposed method.