The invention discloses a graph learning model based on a reconstructed graph, and belongs to the field of image annotation, and the method comprises the following steps: searching a semantic nearestneighbor of a test image through an improved nearest neighbor algorithm, constructing a similar matrix, carrying out the clustering of the image through a random dot product graph, mining the internalconnection of the image, obtaining a weighted similar matrix, and obtaining a preliminary image annotation result through a graph learning algorithm. The relation between the labels is used for labeling, the co-occurrence imbalance between the labels is considered in the process, a nearest graph theory model is introduced, and the problem of label imbalance is effectively solved. The random dot product image is used for reconstructing a transfer matrix of labels, and the problem of image label coexistence asymmetry is solved. Further, a Naive Bayes nearest neighbor classifier is used to establish a joint likelihood function between the image and the tag. The image labeling model based on the reconstructed image model is provided according to the characteristic of unbalanced classificationof the image labels, and the recall rate of the labels can be effectively increased.