I need a fast approach for object localization in kind of images used in the industry, which are related to different parts of car and the goal is to categorize them in different classes. At the first step, I want to detect the part in image and remove the background for making the rest of the process easier and faster. It shouldn't necessarily extract just the object and there is no problem if some part of the background is extracted as well, as long as the line that separate the detected part doesn't have any texture. for example it can be a rectangle or circle.
As this process should be done when an object (part of car) is crossing a conveyor belt, it must be a really fast process.
Thank you so much.
1. Capture the empty (only background) screen
2. Capture your image
3. Substract last image from backgroung image
4. Threshold
Result is your objects
as you know segmentation is the first step in image analyse. i think otsu thresholding method can be helpful.
for the detection I suggest the follwoing:
1. If you know information about the background can can build a model ( for example by taking several images of empty background) all what you need is to substract every image from background one and do some morphological operations
2. If it is difficult to build a background model, you have tow options, which depends on the complexity of the object features:
a. you can use a template matching if the objects are not very similar and the size and oriantation is not varying so much,
b. if the objects are complex, you may need to train a detector for each category of the objects
If you need to automatically classify the objects that is another problem which may requires to train a muticlass classifier.
For a fixed camera indoors with good lighting, simply differencing consecutive frames is very fast and surprisingly effective for removing the background. This (FIR) filter has a pole at z=0 and a zero at z=1. If this gives too much noise, try looking at a more sophisticated low-pass differentiator (FIR or IIR).
I guess you should use Top-Down saliency approach for segmentation. In this approach, the detection and segmentation of concerned parts/areas/regions is done while rest of background is discarded.
It would help if you showed us a few of your actual images. Depending on how complex they are, different techinques might be the answer. Also, please tell us how stable the lighting conditions are.
Actually, it's the first step for classification in my project. Now I use a method for separating the object in this way. I marked some dots around the object before capturing the image and then used them to apply a convex hull around it.
Actually, the main goal is to end up capturing some of the background with a simple and fast approach.
In fact the lighting is not consider in the process and is totally stable.
1. Capture the empty (only background) screen
2. Capture your image
3. Substract last image from backgroung image
4. Threshold
Result is your objects
SIOX method could interest you a lot
http://www.siox.org/
http://fiji.sc/SIOX:_Simple_Interactive_Object_Extraction
http://en.wikipedia.org/wiki/Simple_interactive_object_extraction
It is an open-source approach, it is implemented in ImageJ and Fiji and is very well known in the imageJ community. We discovered it 6 months ago and we were amazed by its easyness to use and its processing speed.
Check the links to have a good idea of the method itself.
Good luck !
As others have said, background subtraction is quite suitable and very quick to implement. Your sample image suggests you might have a problem with shadows though.
Something like this might help on that aspect:
Single-image shadow detection and removal using paired regions
R Guo, Q Dai, D Hoiem - Computer Vision and Pattern Recognition, 2011
An other interesting paper, very powerful method but maybe not so easy to implement :
"Contour and Texture Analysis for Image Segmentation" , Malik et al., International Journal of Computer Vision 43(1), 7-27, 2001
This subject is not in my area of specialization. I recommend to contact Dr. Shams Esfand Abadi at [email protected]
Dear Marjan Ramin, when marked some dots around the object before capturing the image, you spend more additional time. The adequate method in this case is segmentation with thresholding or HPF, then subtracting the background.
I think the fastest but not necessarily the most accurate is using color information for segmentation, or using simple threshold (local or global) on the histogram of the image.
you may learn the color information or the threshold offline (for the objects of interest), then use them directly online in segmentation because you will have the same camera and circumstances.
g(x,y)=1 if f(x,y) >threshold, 0 otherwise.
now it is easy to determine the area enclosed your object, perhaps, you will get the exact boundary of your object.
use those info to extract you object to be classified in a later step.
Background image can be obtained by capturing from the camera in the condition no object in front of the camera. After that you can do image subtraction as mentioned by Mahmut Sinecen.
Otherwise you can use Gaussian to model the background and then do the image subtraction.
Did I you still need help with this? We are about to release a very powerful and intuitive 2D/3Dimage analysis package at MIPAR.us and I'd be happy to try an image to see if we an help you. MIPAR has been to tack image segmentation and analysis tasks for us that few other software could.
- Badges
- Science topic