Automatic Recognition of Weather Records

Even before the use of computers, meteorological data was recorded for statistics by means of handwritten notes. In addition to those handwritten documents stored in historic archives, processes, as manually filling in measurement values into forms, still depend on pen and paper. However, manually digitizing this data for further processing is cumbersome. In this report, a character recognition system for automatically digitizing handwritten weather records is proposed. The weather record dataset regarded in this report consists of known printed forms with handwritten meteorological measurements. The scope of this report are the numeric temperature values which are measured at three different points of time. The localization of the numerical data is achieved by first reconstructing the tabular structure of the form. Using vertical and horizontal projection profiles, the rough positions of the lines building up the table are found. Errors in the layout analysis are corrected using a-propri information of the form. Additionally, to reduce the influence of spurious lines, a stroke preserving line removal method is proposed. The extraction of the digits and signs is done using a binarization based on the Savakis filter. Using a trained PCA basis as a filterbank the features of the digits and signs are extracted and subsequently classified with multiple SVM with RBF kernels. Moreover, to allow an identification of uncertain prediction results, the class probabilities are estimated. The evaluation was conducted using three different digit databases with manually annotated ground truth, synthetically generated digit images with spurious lines and weather records from five different measurement stations. On a dataset with weather records from five different measurement stations an accuracy of 93% per digit is achieved. Furthermore, on a dataset containing only weather records from a single writer the performance is improved to over 99%.