Research On Engineering Drawing Vectorization Post-Processing Technique Based On Recognition Quality Evaluation

Engineering drawing recognition technique aims at converting scanned engineering drawing images into vector formats compatible with popular CAD systems. Getting a vectorization result with sufficient accuracy is the basis for further drawing interpretation. Unfortunately, none drawing conversion systems can achieve the satisfactory recognizing accuracy. Correcting manually for vectorized errors is needed and it is a time-consuming process. Research on improving output data qualities from vectorization system is significant for enhancing performance of raster-to-vector systems. In this dissertation, the author presents a recognition quality evaluating mechanism that provides the system with the ability of measure the quality of each primitive generated. .One of the focuses in this dissertation is the design of a recognition quality evaluator for graph entities. The original image information corresponding to the graph entities is considered as a constrained template or allowable error range. Some evaluating factors, defined by using fuzzy theory, have been proposed to express a graph entity's recognition quality and as the input of the quality evaluator.To get the evaluating result, a recognition quality evaluator for straight lines based on artificial neural networks has been designed. After training the neural network can work well and output the quality evaluating results. Furthermore, this proposed evaluating approach is not sensitive to scanning image resolution. For the graph entities with arc type, we mainly studied a mis-recognition in graph entity types, referred as "pseudo arc". By searching the corner points on the border of the arc and divide the arc to piece-wise, the analysis is then done for every piece respectively.For those graph entities with poor recognized results indicated by the quality evaluator, further processing operations for them have to be considered. A correction idea based on re-vectorization has been put forward. Re-vectorization is performed on the basis of original vectorization by obtaining the initial parameters such as line types and slope. Inuring tricing it can adjustthe parameter adaptively. To decrease the influence of noise pixels, we designed a stop probe to detect tracing-stop conditions. Process of re-vectorization is performed recursively until quality of graph entity is satisfactory.Finally, some key technologies for 2D graphs reconstruction are studied. The issues concerned include the recognition and presentation of topologic relations among graphs entities and a directed and weighted graph model established to describe the relations and techniques on dimension driven. The recognition for assistant type graph entities in engineering drawings, such as dash-dot lines, dash lines and hatched section lines, is investigated.