JPS63298669A - Recognizing system for drawing - Google Patents

Abstract

PURPOSE:To improve recognition efficiency by extracting candidates of interpretation for each data unit, calculating the certainty CF (Certainty Factor) of the candidates by using a correspondence rule and a possibility distribution table, extracting basis patterns, and discarding candidates which are not extracted as the basic patterns. CONSTITUTION:A data unit (processing unit) of each level is determined. Then candidates of interpretation are extracted for each data unit and the certainty CF of the candidates is calculated by using the correspondence rule which make candidate categories of low-order and high-order levels correspond to each other and the possibility distribution table where the level of the possibility of correspondence is digitized. This operation is carried out for each level from the low-order level to the high-order level and candidates with CF values less than a specific threshold value are discarded; and basic patterns are extracted for a candidate group which is remained up to the highest-order level and candidates which are not extracted as the basic patterns are discarded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to automatic recognition of two-dimensional mechanical design drawings for sheet metal processing, etc., and a drawing recognition method that can be used as an input means to CAD.

In recent years, much research has been conducted into using computers to automatically recognize handwritten mechanical design drawings. In particular, attempts have been made to introduce knowledge processing and handle various types of ambiguity.

Research on the interpretation of drawings in such automatic drawing recognition includes the following.

(1), Hatakeyama, Yuyo (Toshiba): “Drawing reading device TO
AGRAPH ~ Application to parts drawings (1).

(2) ~5S-2.58-3” Information Society of Japan, National Conference,
Showa 61. Late period.

(2) Hine, et al. (Matsushita): “Automatic mechanical drawing recognition system ARCADIA-M” Information Society of Japan, National Conference, Showa,
61. Late stage, 4N-7, 4N-8, 4N-9, 4N-
10゜(3), Ito et al. (Tokyo Institute of Technology) “Interpretation of mechanical drawings using Brang networks and geometric models” PRL
84-51゜(4), Morimoto, Ariki, et al. (Kyoto University) “Drawing structural analysis using drawing composition knowledge” PRU86-12゜However,
The above conventional technology has the following drawbacks. That is,
In (1) and (2) above, the drawing element analysis part is a procedural process, and it is difficult to easily deal with variations in entry conditions. Moreover, when handling ambiguous information during recognition, or when interpretation cannot be done uniquely (
The handling of ambiguity of interpretation) is not considered. the above(
Regarding 3), it is also a procedural process, and vague knowledge (
The disadvantage is that the information and rules) are not handled adequately. Regarding these studies (1) to (3),
In (4) above, multiple interpretations are allowed, and the reliability (certa
ambiguous knowledge is handled by introducing inty factor (hereinafter referred to as CF). However, this method of correcting the CF value is entirely empirical and has not been formulated.

1-11 In line with the above-mentioned research trends, the present invention flexibly deals with changes in drawing entry conditions by introducing knowledge processing, and eliminates various ambiguities (uncertainties in information, knowledge, etc.). The purpose of this work is to provide a drawing recognition method that improves recognition efficiency by formulating the handling of ambiguity of interpretation.

In order to achieve the above object, the present invention defines a processing hierarchy with different candidate categories in a drawing recognition method in an automatic drawing recognition system that automatically recognizes drawings using a computer, The data units (processing units) at each level are determined through the integration process, candidates for interpretation are listed for each data unit, and the probability CF (certainty factor) of the candidates is calculated.
) is calculated using correspondence rules representing rules for associating lower-level candidate categories and higher-level candidate categories, and a possibility distribution table that quantifies the strength of the possibility of correspondence, These operations are performed for each level from the lower level to the upper level, candidates whose CF value is less than a predetermined threshold are rejected, and basic patterns are extracted from the remaining candidates up to the highest level. This method is characterized in that drawings are recognized by rejecting candidates that are not extracted as such. below,
An explanation will be given based on an example of the present invention.

FIG. 1 shows the overall processing flow of an automatic mechanical drawing recognition system, which is the premise of the present invention.

The present invention proposes an algorithm for "diagram interpretation" of 5TEP 5 in the processing flow shown in FIG.

FIG. 2 shows the processing hierarchy for drawing recognition processing based on the present invention using a design drawing for sheet metal processing as an example.
The interpretation of the drawing follows this processing hierarchy from bottom to top, ie, from level 1 to level 5.

FIG. 3 shows the processing flow for interpreting the diagram.

The data units (processing units) at each level are determined through the integration process, candidates for interpretation are listed for each data unit, and the probability CF (c
ertainty factor). As the CF, "basic probability" as defined in Depster & Hafer theory (hereinafter referred to as D&S theory) is employed. This calculation of basic probability uses a "correspondence rule" and a "probability distribution table.""Correspondencerules" represent rules for associating lower-level candidate categories with higher-level candidate categories, and are
It takes the form of an F-THEN rule. The IF section describes preconditions that are necessary for association and that must be satisfied by the data unit. Fuzzy expressions are allowed in the description of this precondition. In addition, the "possibility distribution table" quantifies the strength of the corresponding possibility (0 to 100%
) and are set for each rule. The basic probability calculated from each rule is the weighted average value or
It is synthesized using the D&S theory synthesis method. This operation is performed for each level from level 3 to level 4 to level 5, and candidates whose basic probabilities are below a certain threshold along the way are considered to have low reliability and are rejected. In this way, basic patterns are extracted from the candidate group remaining up to level 5 (as a general rule, only the elements (interpretations) that make up the basic pattern are selected as final candidates), and candidates that are not extracted as basic patterns are rejected. Ru.

The above is an outline of the present invention, and each process will be explained in detail below with reference to FIGS. 4 to 8.

FIG. 4 is a diagram showing a processing flow at level 3, and a connected segment in this processing is a data unit in which 8 black pixels are connected. Further, a short segment is a data unit generated by performing segmentation (division by feature points such as branch points) on a pixel column. Level 3 data units are extracted and integrated from these two segment lists. Furthermore, the feature amount as an image is calculated, and the basic probability is calculated by referring to the possibility distribution table. This calculation method is [Bungou (Ricoh): “D
"Handling Uncertainty Using Empster and 5hafer's Theory" Technical Memo D A H22-13 ('86.
10.7)] (hereinafter referred to as Document 1). A plurality of feature quantities are considered, and each basic probability is synthesized using the synthesis side of the D&S theory. The following can be considered as the special quantity.

・Size of the circumscribed rectangle ・Extensibility of the circumscribed rectangle ・Isolation ・Line density ・Black pixel density ・Is it an intersection of 3 straight lines that looks like an arrow? , is the outline triangular? ) ・Is the outline circular? ) Figures 5 and 6 show the flow of the inference method in drawing recognition of the present invention (Figure 6 (a) shows the data unit, and (b) shows the basic probability), and the "inference"
means to interpret elements lower in the processing hierarchy into elements above. is defined by generating a dataset of A data unit, which is a unit at a certain level in the processing hierarchy, has a larger spatial extent as the level is higher. Therefore, each time you go up one level in the hierarchy, it is necessary to integrate data from several lower levels and recreate a new set of data units that serve as processing units. However, later on,
It is necessary to calculate the basic probability for the candidate set of each data unit (FIG. 6(b)). These processes are performed from level 3 to level 5.

As the above-mentioned integration processing, the following Tables 1 to 3 can be considered.

Table 3 (integration processing at level 5) The processing unit is determined by the integration process, but next.

The subsequent calculation method of the CF value (basic probability) will be explained using FIGS. 7 and 8. In FIG. 7, the upper level is called level β, and the lower level is called level α (=β-1). After the integration processing at level β (1 in Figure 7), register the data unit that will be the processing unit (2 in Figure 7),
After performing this process for all data units, the candidate category and its CF are added to each data unit.
Calculate the value.

Let Bk be the data unit you want to calculate (3 in Figure 7)
, Bk, one of the lower level α data units is Ak2 (Q=1...Ω.). Generally, there are multiple candidate categories for Akfi, among which the corresponding rule group (Qs) ( S = 1...NR) The essential conditions for the upper level candidate category bj (j = 1...m) are the rules written in the IF part, and the prerequisite part Qs to some extent, although not essential. Bk obtained as a result of matching bJ with the rule that holds (these two rule groups are R) is calculated (5 in FIG. 7).

Regarding the candidate species i and the rule S(εR) at these values α, there are two first-order methods (1) and (2) regarding the algorithm.

However, c (ai) = m', QCae, ) P (q,) - weight kQ q; prerequisite for Q S R (rule set); essential rule + rule that is not essential but holds premise 1 to some extent ■: [)Hpster synthesis Below, the above methods 〇 and ■ will be explained in order.

1 dog■ This method is a type of direct method of fuzzy inference, and was published by Kanno (Tokyo Institute of Technology): ``Fuzzy Control'' 2nd Fuzzy Systems Symposium ('86.6.16-18) Pi-p
8. In general, fuzzy modus pone is similar to "Inference method-3" in fuzzy control by
Fuzzy reasoning by ns, i.e. CAz-A1→B1)=>B.

In the direct method of (A (Bl: fuzzy expression defined on set

(Tsukamoto (Tokyo Institute of Technology) 'Fuzzy Reasoning' Measurement and Control Vol. 1.22 No. 1 (558, 1) p. 139-p
145), and the 2nd Fuzzy System Symposium on "Fuzzy Control" by Sugano (Tokyo Institute of Technology) ('86.6.
16-18) In "Inference Method-3" of pi-P8, the nonlinear ambiguous relationship R is approximated as follows. Divide the input spaces X and Y into subspaces i, e, A1, respectively.
．． B□ is divided into several fuzzy sets), and the fuzzy relation R,:A −+B in that subspace is expressed by the x Qifli function f& of the consequent part of each IF-THEN rule. Furthermore, the synthesis side “O” of the inference is expressed as “weighted Ωi” by the satisfaction level P(A,) of the antecedent part A.
In the 0-line processing method that approximates by "average", the probability distribution table corresponds to the function f>. However, when taking the weighted average, normalization is also performed based on the number of corresponding rules to be considered. Also, if several lower-level data units are integrated for one data unit, the basic probability must be determined by an appropriate calculation, but here we will take a simple average. (Step 8 in Figure 7).

j 0m<b)=%mk″<bj)/n.

β k l β k 1 Inudo Let Prop(a) be the proposition that the lower level data unit targeted for processing is the candidate category assumed in the antecedent part of the correspondence rule Qs, then the method ■
Now, instead of just using the probability distribution function of the correspondence rule Qs to calculate the basic probability.

As a weight factor for evidence, P (Prop(a)△
Calculate the basic probability by multiplying A , )s Ql . To compose the basic probabilities,
Using the join side of Dempster & 5hafer,
Also.

Perform normalization based on the number of correspondence rules to be considered and a method of combining basic probabilities between data units (step 8 in Figure 7).
).

In this way, Bk performs this calculation for candidate Bk according to method ■ or ■.As an interpretation candidate for each Bk, "B
The “root hypothesis” that “k is C1” and “Bk is C
There are two types of ``disjunctive compound board theory'': ``C or C2... or Cj (j≦m)'', but the basic probability assigned to the compound hypothesis in step 10 of Figure 7 CF value) among the root hypotheses (for example, C
equally divided between j).

The structure of the possibility distribution table used in steps 5 and 6 in FIG. 7 will be explained below.

Table 4 shows an example of a possibility distribution table, which summarizes the strength of the possibility of correspondence between categories belonging to different processing hierarchies. By defining each corresponding rule, it is also possible to express the priority between rules. Furthermore, when calculating the basic probability from the possibility distribution table, the method proposed in the above-mentioned document (1) is followed.

Table 4 Examples of corresponding rules are shown below.

In the following, we refer to the upper level data unit as U.

The lower level data unit sequence that constitutes it is (U(
i)) (i = 1...n, number adjacent data units in order), the data unit of interest is U(i), and rules with non-essential preconditions are marked with an “x”.

Example of correspondence rule from level 3 to level 4 (i) If the candidate category Dj of the corresponding light is noise, then noise...n=1 1F...U(i) seems to be a letter AND U(i ) is isolated (no other data exists nearby) THE
N...U may be noise IF...U (
i) seems to be a dotted line A N D U (i) is isolated THEN...U may be noise IF...
・A N D U (i) where U (i) seems to be a short straight line
is isolated THEN...U may be noise IF...
・U (i) is a short curve element ANDU (i) is isolated THEN...U may be noise (n)D
If , is a character/symbol area, × Engineering F...U (i) seems to be a character A N D U
(ill) seems to be a character. THEN U is very likely to be a character area (
C1,,,n-1) ×IF...U(i) seems to be a letter A N D U (
ill) seems to be a line element (LL, LS, KL, KS, DL) ANDU (ill) seems to be literature THEN...U may be a character area (i=
1. ,, n-2) Figure 9 is a diagram showing an example of a basic pattern, and the pattern types in Figure 9 correspond to Table 5 (No. in Table 5, No. in Figure 9 in part 1, It is easier to understand if you match the 1st part with No. in Table 5 and the 2nd part with No. 002 in Figure 9).

The relationships between constituent elements are mainly expressed using binary predicates, and are expressed concisely. The basic pattern mainly describes how to enter dimensions, but the outline line is also considered a kind of basic pattern with the condition of ``creating a closed loop'' and is extracted at this stage. - Table 5 Figure 10 shows the overall data structure, where ■ is a data table, ■ is a correspondence table, ■ is a basic pattern instance, ■ is a feature point list, ■ is original image data, and ■ is a coordinate point list. Level data is associated with each level using a ``correspondence table''. In addition, the connection relationship between data units within each level is expressed as a “feature point list”.
It is described in

The data structure of each level is shown in the table below.

Table 6 (Composition of coordinate point list) Data table example (level 2) Table 7 (short segment) Table 8 (consolidated segments) Data table example (level 3) however, LL: long line element LS: Short straight ahead KL: long oil element KS: short curve element AA: Arrow CC: Characters, symbols BB: Black circle Table 10 (data structure of correspondence table) Table 11 (data structure of feature point list) The specific example described above has the following advantages.

1. In the processing at level 3 shown in Figure 3, only the large connected segments are thinned, and 4.5 in Figure 4). Extract various feature quantities using D&S
Since it uses theory to assign basic probabilities to candidates, it is faster because characters belong to small connected segments and do not need to be thinned.

Furthermore, since black circles and arrows are difficult to extract from a thin line image, it is effective to use connected segments. The D&S theory is effective in determining the category of a data unit from various feature values.

2. Performing top-down integration processing (Table 1.2
．． 3) Processing units that are more spatially expansive (
For example, it is effective for finding dashed circles).

3. As in the process of mapping (interpretation) between hierarchies shown in Figures 7 and 8, when determining the category of a certain data unit, use the information of the lower-level data units that make up the data unit. This formulation not only increases the judgment ability but also increases the reliability of the judgment results.

Normalizing by the number of rules N when taking a weighted average, as in equation (2), will not lead to unreasonable results even if rules are added or deleted.

Also, when normalizing, by counting "rules that are not essential but whose premises hold true to some extent" in the number of rules NR, it is possible to handle any number of such rules (rules that are not essential) without a limit on the number. I have to.

Similarly, taking normalization by N into consideration during synthesis based on the D&S theory is similar to the effect described in 4 above. Also, D&
By using S theory, the correspondence to the correct answer candidate (C
F value calculation) ability can be improved.

Taking a simple average requires less calculation and can speed up processing.

According to the D&S theory, the same routines can be used and the amount of program memory can be reduced. In addition, normalizing by the number of components also reduces the recognition results at lower levels (
The dependence of CF value calculation on how many parts are recognized can be reduced.

8. By having a probability distribution table shown in Table 4 for each corresponding rule, the priority between rules can be expressed.

Contributes to improving recognition rate.

9. The correspondence rules are made easier to understand by using IF-THEN rules that allow fuzzy expressions and have a precondition part that combines the declaration part of the candidate category that is the premise and other precondition parts by conjunction. This makes it easier to write down images, and it also makes it possible to deal with recognition errors (due to noise, blurring, etc.) at the image recognition level.

10. By creating a data table and feature point list for each level and managing data correspondence between levels using the correspondence table, it becomes easy to display, search, and modify data units at that level. Additionally, these can be seen as representing the process of interpretation, and can be used to display the cognitive process.

In addition, the present invention extends from the automatic recognition of two-dimensional mechanical design drawings as in the above embodiment to the automatic recognition of three-dimensional mechanical design drawings (3-dimensional mechanical design drawings).
It can also be applied to automatic recognition of drawings) and general drawings (maps, logical circuit diagrams, etc.).

Effects As is clear from the above explanation, according to the present invention, the handling of uncertainty in knowledge (recognition information, rules regarding drawings) is formulated, and by allowing ambiguity (multiple interpretations) of interpretation, one variety can be solved. You can use the knowledge of

Can increase cognitive ability. In addition, the hierarchical processing makes it easier to write down corresponding rules (rules are easier to classify), and it is also easier to deal with changes in entry conditions. In addition, the overall process flow is simple, such as repeating the integration process and the basic probability calculation several times, making it easy to implement computationally. Furthermore, in a bottom-up search that allows multiple candidates, the number of search trees increases explosively, but in the present invention, a basic pattern is defined and top-level candidates that do not constitute the basic pattern are rejected. Since it is a pruning method, there is no explosive increase in the number of search trees, and it is possible to increase processing speed and save memory.

[Brief explanation of drawings]

Fig. 1 is a diagram showing the overall processing flow of the automatic mechanical drawing recognition system, Fig. 2 is a diagram showing an example of a processing hierarchy in the drawing recognition process according to the present invention, and Fig. 3 is a diagram showing the processing flow of the drawing recognition process. Figure 4 is a diagram showing the processing flow at level 3 in the processing flow shown in Figure 3, Figure 5 is a flow diagram of the inference method in drawing recognition, and Figure 6 is the integrated processing in drawing recognition of the present invention. FIG. 7 is a process flow diagram for calculating the basic probability in the drawing recognition method of the present invention, FIG. 8 is an explanatory diagram for calculating the basic probability, and FIG. 9 is a basic pattern in the drawing recognition method of the present invention. FIG. 10 is a diagram showing the overall data structure in the drawing recognition system of the present invention. ■...Data table, ■...Correspondence table, ■...Basic pattern instance, ■...Feature point list, ■...
・・Original image data, ■・・Coordinate point list. Figure 1 Misfortune 2 rA Figure 3 Figure 414 Figure 5 Figure 6 (bnmαI m(<2 Figure 7 Figure 9 Figure 10

Claims (1)

[Claims] In the drawing recognition method of an automatic drawing recognition system that automatically recognizes drawings using a computer, a processing hierarchy is defined for each different candidate category, and a data unit (processing unit) at each level is determined through integrated processing. For each data unit, candidates for interpretation are listed and the certainty of the candidate is determined by CF (certainty).
factor) is calculated using correspondence rules that express rules for associating lower-level candidate categories and higher-level candidate categories, and a possibility distribution table that quantifies the strength of the possibility of correspondence. , These operations are performed for each level from the lower level to the upper level, candidates whose CF value is less than a predetermined threshold are rejected, and basic patterns are extracted from the remaining candidate groups up to the highest level. A drawing recognition method characterized by recognizing a drawing by rejecting candidates that are not extracted as a pattern.