The invention discloses an image retrieval method based on a self-adaptive rectangular window. The method specifically comprises the following steps: step A, extracting local unchanged characteristics of database images, clustering space coordinates of characteristic descriptors in each database image using a G-means method; step B, according to classification of G-means clustering in each database image, establishing a self-adaptive rectangular window, and removing a sparse rectangular window, merging rectangular windows, and removing small rectangular windows; step C, respectively vectorizing the self-adaptive rectangular windows in the database images, establishing self-adaptive window vectors, and based on the window vectors, establishing an inverted index; step D, vectorizing a to-be-retrieved target region, performing similarity search in the inverted index, to giving a final retrieval result. Compared with the prior art, the method improves retrieval accuracy while ensuring retrieval efficiency.