Analysis of Dimension in Mechanical Engineering Drawings

Many paper engineering drawings are still in use and there is a demand for reliable means of transforming them into digital information and incorporating them into a CAD (Computer Aided Design) system. To be efficiently used, paper engineering drawings dimensions must be analysed allowing for example 3D model interpretation. We describe the two majors parts of our system: vectorization and dimension identification. The first part, vectorization, uses a specific algorithm to recognize as soon as possible, text and arrows. This is a typical case where a priori knowledge can be introduced w well in low-level processing as a t higher-level. The second part is achieved in the same way, finding as soon as possible some acceptable dimension. Two different~ actors are introduce two implement the dimension identification: the "Grouper", which groups elements, trying to build dimensions, and the "Checker" which controls the validity of the set of grouped elements. These two actors are working alternately and share a common view of the drawing. The lS' low-level part, is already implemented in C. The 2nd, the high-level one is implemented in Objective-C, which seems powerful for handling the complexity of the system. We conclude with remainings problems and furthers works. 1 Vectorization and Symbols Extraction. The vectorization is the start point of the analysis system, and the whole interpretation depends of the precision of the extracted symbols. The vectorization must be as complete as possible. The basic vectorization is processed by the REDRAW system [I], which provide a set of vectors and a connected component tree including the relations between them. The advantage of this structure is that immediate access to topological relations and to pixels is always possible [11,3] Components are also defined with their contours. However, this vectorization is not sufficient to allow the analysis of the dimensions. We must add more specific processing stage for taking into account the characteristic entities of the dimensions, which are arrows, text and thin lines. A good recognition of these elements are essential for the detection of the entire dimension. Thin lines can be easily separated from the others by consideration of thickness, computed in the vectorization step, but arrows and text can be seen = composite symbols, and need a specific process [I]. Text is extracted from graphic by analysing the connected component structure, area and density (81, and further, characters are grouped into strings. This method do not require recognition of individual character, and work in any direction, even slanted one. Often, arrows are very small symbols and their recognition is essential to go on the analysis. To avoid big and irreversible distortions, their detection must take place before vectorization, just after polygonal approximation. A lot of methods are known to solve this classical pattern matching problem: subgraph isomorphism [9,12], Hough transform, or looking for main domain features, in this case a triangular form. This is the method we have chosen [I]. The result is a set of possible arrows. Some of them are not arrow, but this is not a serious problem, because the next stage will detect them as bad ones. Figure 1 shows the results of textlgraphic separation and arrows detection. 2 Dimension Analysis and Drawing Association. Localization and recognition of dimensions is an essential contribution to a complete understanding of the whole drawing. We use two kinds of knowledge representation: knowledge represented by rules and knowledge represented by grammar. Different elements can be a part of a dimension: text, arrow, shape, witness, tail and contour. Figure 2 shows these different parts. Figure 1: Arrows detection and Text/Graphic separation. Recognized arrows and texts are boxed on this figure. Figure 2: Different parts form a dimension: text, arrow, shape, witness, tail and contour. The graphical representation used in technical documents are most of time very strongly structured, following a set of well known rules. Our idea is to use as much as possible these rules to lead the interpretation. Dori [6,7] has studied recognition of dimensions; his method consists of a ayntactical analysis of the dimensions. A grammar describes all possitde dimension types and there is a subgrammar for text recognition. The grammar's rules enable the recognition of any dimension in a mechanical drawing. The checking part of our work is based on his method. On the other side, we use domain knowledge to build gradually dimensions. Our global strategy is first, t o find ordinary dimensions, and later, to find more complex ones. I t is important to provide as quickly as possible high-level entities for allowing to be more selective in the next choices. To achieve this strategy, we have introduced two "actors": the Grouper which try to build dimensions, and the Chec. ker which controls the validity of the grouped elements. A permanent dialogue is set up between the two actors, until a final solution is founded. The two actors are working alternately, and share a common view of the drawing. As the work is going on, their behaviour is changing with respect to previous success or fails in the grouping or recognition steps. They become more and more selective as the system is going on. Backtracking to previous data, pixels data for example, is even possible. Next two subsections give details about the Grouper and the Checker idea.