The invention discloses a CRNN-based picture table extraction method, and the method comprises the following steps: S1, carrying out the perspective transformation of a to-be-detected picture, and correcting the picture; S2, performing table skeleton extraction on the corrected picture by using a deep neural network; S3, obtaining a cell ROI (Region of Interest) from the table skeleton; S4, recognizing text contents in all the cell ROIs through an OCR recognition model; and S5, restoring the text content to a table through the table skeleton typesetting in the step S2, so that the picture table is converted into a data table, and the extraction of the picture table is completed. According to the method, perspective transformation is carried out on the to-be-recognized picture once, the picture angle is corrected, and then the deep neural network model is used for extracting the overall table skeleton, so that the situation that the edges of cells are connected by handwritten characters in the prior art or the problems that a table in the picture is not clear, the picture is light, inclined and fuzzy are solved, and the problem that time and labor are wasted because a large amount of manual parameter adjustment is needed is solved.