The invention discloses a convolutional neural network-based character recognition method, and mainly aims at solving the problem that existing manual engineering drawing data input is time-consuming,labor-consuming and high in error rate. The method comprises the following main steps of: 1) converting grayscale images into binary images by adoption of Otsu method-based optimal global threshold value processing; 2) rotating to correct inclined images; 3) extracting a table box by adoption of a mathematic morphological algorithm; 4) obtaining an area where characters are located; 5) rotating inclined character strings for a corresponding angle to a horizontal direction; 6) searching circles in the images by adoption of a circular Hough transformation algorithm, and extracting welding beadnumbers in an engineering drawing; 7) carrying out image-text recognition on the engineering drawing by adoption of a convolutional neural network method; 8) correcting simple errors through an encoding rule and automatically correcting recognition errors; and 9) outputting and storing the drawing data. In image-text recognition of engineering drawings, the method has high correctness and timeliness, and is capable of realizing efficient recording and management of engineering drawing data.