The developed software system is aimed at visual inspection of motorcycle connecting rods. The system is able to analyse the dimensions of two different types of connecting rods to allow a vision-guided robot to pick and sort rods based on their type and dimensions. The two rod types are characterized by a different number of holes: Type A rods have one hole whilst Type B rods have two holes. Below it has been illustrated a few examples of the pictures which the software has been programmed to analyze; the full list can be foud here.
The developed software is capable of acquiring the following features from each image:
- Number of Rods
- Type of Rod (Type A or Type B)
- Position and Orientation
- Length, Width, Width at the barycenter
- For each hole: position of the barycenter and diameter
The program is robust to the following possible issues one might run into when analyzing the rods:
- Iron Powder hindering segmentation
- Rods touching with one another
- Images might contain objects which are not rods and should not be further analyzed
The picture above depicts the whole project's workflow.
The first step consists in the foregorund/background segmentation of the image, which has been implemented with the Otsu's algorithm. This was possible because the images have been acquired through the so called backlighting technique, hence their histogram appear inherently binary. However, some of the images are affected by the presence of iron powder, which must be smoothed out before the segmentation phase otherwise they might be wrongly classified as foreground pixels. This issue has been tackled through the implementation of a median filter which smooths the image from the iron powder which is treated as impulse noise.
Once the image has been correctly segmentated every element belonging to the foreground must be told apart with the aid of the Connected-Component Labeling. Before this can be performed, however, every touching rod must be separated otherwise two separate rods might be labeled as belonging to one single rod. In order to detach touching components each component have been enclosed with a convex hull and then the points where the rods are touching are computed through the convexity defect function.
Below it is demonstrated how two touching rods are separated.
Each rod is assigned a label in the Connected-Component labeling and a RGB-mask is applied to highlight them.
Before performing the Blob Analysis where the features of each Blob is computed, those elements present in the pictures which are not rods(washers and screws) are discarded.
The results of the blob analysis are depicted below:
The results are also printed on screen and saved in an excel file.
[ Rod type: A
num holes: 1 -> holes: [{'D': 24.70020388546793, 'Cx': 45.06264775413712, 'Cy': 115.52009456264776}]
ib: 82.35443425076453, jb: 82.12721712538226
orientation: 47.951162391517 8barycenter width: 16.278820596099706
width: 44.72901342652703 length: 151.86274093262938
, Rod type: B
num holes: 2 -> holes: [{'D': 23.799887584719077, 'Cx': 86.9047619047619, 'Cy': 119.10273368606701}, {'D': 26.68729810895878, 'Cx': 163.10987529491067, 'Cy': 114.71149309066396}]
ib: 128.38539553752537, jb: 116.62407031778228
orientation: 86.69539973760394 barycenter width: 17.029386365926403
width: 35.345310989417506 length: 100.46780517663012
, Rod type: A
num holes: 1 -> holes: [{'D': 25.150362035842353, 'Cx': 149.78933333333333, 'Cy': 167.61904761904762}]
ib: 104.70156106519742, jb: 172.1900826446281
orientation: 84.62185250686889 barycenter width: 19.1049731745428
width: 50.06107504951664 length: 136.99468496705683
]Check the documentation for more in-depth information.
Clone the project
git clone https://github.com/CDOrtona/Rod_Inspection_Machine_Vision.gitGo to the project directory
cd Rod_Inspection_Machine_VisionInstall the required python modules:
pip install -r requirements.txtDistributed under the MIT License. See LICENSE.txt for more information.