Clustering algorithm for recognition of computer aided design images

Tuyển tập Hội nghị Khoa học thường niên năm 2019 ISBN 978 604 82 2981 8 180 CLUSTERING ALGORITHM FOR RECOGNITION OF COMPUTER AIDED DESIGN IMAGES Nguyễn Văn Nam Thuyloi University 1 INTRODUCTION Comput[.]

CLUSTERING ALGORITHM FOR RECOGNITION OF

COMPUTER AIDED DESIGN IMAGES

Nguyễn Văn Nam

Thuyloi University

1 INTRODUCTION

Computer Aided Design (CAD) images are

technical drawings that are frequently used in

engineering domains In mechanical factories,

CAD drawings must be converted to the

corresponding Computer Numerical Control

(CNC) machine commands for cutting material

Most CAD software can automatically

effectuate this conversion for CAD files

However, the transformation of scanned CAD

images are currently done by human beings

Most recent recognition methods based on deep

neural networks are not efficient enough since

the CAD objects are too small and noisy In this

paper, we rely on old but efficient DBSCAN

clustering algorithm which produces more than

90% accuracy for cad recognition To the best

of our knowledge, this is the first to address

this problem

Figure 1 Real 3D object and its four 2D

projections

Given only four 2D projections of a 3D object as in Fig 1, an engineer needs to recognize the details of the CAD drawings including projections (front, rear, left, right, top, down), line (solid, dotted, distance), circle (semi-circle, disk), arcs, text boxes,

He then draws a flattening image corresponding to the real object as can be seen in Fig.2 (below) This last CAD file can then be converted to CNC commands

Figure 2 Flattening CAD image

2 RESEARCH METHOD

Figure 3 Scanned CAD Drawing

As in Fig 3, a cad drawing includes a

rectangle bounding box, a section of notes

which is placed at a side of the bounding box,

a text box describing the material type of the

object, several cad projections demonstrating

the detail shape, size of the object viewed

from at most six directions: top, down, left,

right, front, rear

A projection in a CAD drawing consists of

distance lines, alignment lines, textboxes, a

closed contour of the whole object and some

circles, disks, boxes inside the closed

contour The closed contour may be as

simple as rectangles, polygons, circles or any

more complex combinations of them Since

the final CAD image contains only closed

contours, we will describe our method to

extract this contour in every projections and

then link them together

As previous analysis, the cad drawing

images can be clustered in to distinguished

regions This leads us to use some clustering

algorithms in machine learning K-Means

[1] is a partitioning spatial clustering method

which regroups pixels with the nearest mean

K-Means segments data space into Voronoi

cells This method cannot be applied in this

case since the largest rectangle bounding box

will be the only one K-Means partition

Ward [2] is a hierarchical clustering

algorithm This is bottom-up algorithm which

merges small groups in to bigger ones based

on some agglomerative criteria Once more,

the largest rectangle bounds the whole image

so Ward will consider it at the only partition

DBSCAN [3] is a density-based clustering

method This extract low- and high-density

clusters Therefore, it can find clusters of

arbitrary shapes especially closed form

contours provided their points are close

enough to their neighbors DBSCAN defines

core points and outliers The formers must

form a group of at least minPts points which

the distance between one point to its closest

one in the same group is less than eps The

latter are all the remaining points

The DBSCAN algorithm can be seen as in

Fig 4

Figure 4 DBSCAN clustering algorithm

The algorithm starts from any unvisited

point p in the data space and find all points that are reachable to p One-point q is reachable to p if there is a path from p to q

where all points in the path are close enough

(compared to eps) to the previous one If the

number of reachable points are at least

minPts then a new cluster is recorded and its

points are labelled Otherwise, they are outliers

D ={ p}

eps, minPts cluster = []

Any unlabeled p

in D

N

Y

Find all points reachable to p in D based on eps

At least minPts found?

Record new cluster and label its points

N Begin

End

Y

Based on DBSCAN clustering, our cad

recognition method includes the five steps

The first is to extract the outer and notes

partition Secondly, cad projections are

partitioned Third step aims to remove all

distance lines, alignment lines Next, the

largest contour of the object in each

projection will be revealed Finally, all the

largest contours of projections are linked

together to shape the flattening image

3 RESEARCH RESULTS

The method is testified with 20 random

cad drawing images of 300dpi eps and

minPts are chosen as 7 and 10,

correspondingly 90% of the cases produce

accurate results to extract rectangle bounding

box and notes This is because some notes are

placed far from the bounding box and some

are too close to the projections 98% of

projection partitioning are correct Some

errors are due to the fact that small

projections may have links to their bigger

demonstrations After removing all the line

and distance lines, nearly 100% of largest

contour are extracted from projections

Figure 5 The front projection

Figure 8 Final flattenning result image

Figs 5, 6, 7, 8 show the results of our methods for the cad drawing in Fig 1

4 CONCLUSION

In this paper, we target the problem of cad drawing recognition The method is based on DBSCAN clustering algorithm This produces excellent experiment results with more than 90% of clustering accuracy for 20 random cad drawings In the future, we continue with recognition of small CAD items like small circles, disks

5 REFERENCES

k-means clustering algorithm JSTOR: Applied Statistics, 28, 100 108.

Grouping to Optimize an Objective Function”, Journal of the American

[3] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu 1996 A density-based algorithm for discovering clusters a density-based algorithm for discovering clusters in large spatial databases with noise (KDD'96), Evangelos Simoudis, Jiawei Han, and Usama Fayyad (Eds.)

AAAI Press 226-231

Figure 7 The rear and left (right) projections

Figure 6 Top (down) projections