Multi-stage dense reconstruction method and device based on uncertainty

The invention discloses a multi-stage dense reconstruction method and device based on uncertainty. The method comprises the following steps: acquiring a plurality of original images; generating a multi-scale feature map of each original image through a multi-scale feature extraction network based on cavity convolution; according to the multi-scale feature maps of the original image, determining multi-scale feature bodies of a plurality of source images corresponding to each reference image; top scale reference image features and source image features are extracted, homography transformation is carried out to construct a cost body, and regularization is carried out; generating an uncertainty map by using entropy mapping, and performing weighted fusion on the uncertainty map and the cost body to generate a rough depth map of a first scale; and solving depth increments of other scales by using a Gaussian Newton iteration method, adding the depth increments to the rough depth map of the previous scale, and obtain