- New JobEsterlineProgram Manager
I would like to use the Design of Experiment (DoE) in object detection. My target is to detect airplane door. For this case I extract and detect relevant features in the door such as door Window, Door Handle, Text boxes, Door Arrow, Door Frames in the Airplane door. So I would like to perform DoE to find out the influence of this individual features in the final door position identification. I research some literature and find out that DoE consist of 7 steps.
2. Select process variables
3. Select an experimental design
4. Execute the design
5. Check that the data are consistent with the experimental assumptions
6. Analyze and interpret the results
7. Use/present the results (may lead to further runs or DOE’s).
I have more than 3 inputs factors so guess need to use Comparative
Objective Design approach with Randomized block design or Screening
Objective Design with Fractional factorial or Plackett-Burman design. Is that correct?
So is there any examples or some helps for DoE related to my problem(imige processing, features detection)? Any software tools that can be used or some help for my particular problem?
@astronaut71 – Don’t take this the wrong way, but are you sure you are up to doing this? I’m not sure you understand what you are undertaking. There are measurements to be taken, which will call for conducting a measurement system analysis to ensure they are adequate, as well as conducting the experiments and evaluating the data.
What you show as the seven steps isn’t all that needs to be done. As you say, you have more than 3 inputs, so you may need/choose to conduct a screening design first to see if you can reduce these inputs. You may have strictly linear response, or you may have curvature, in which case you will need to choose the correct design type to ensure that the curve terms end up in the resulting model.
I would encourage you to find a mentor who is familiar with conducting DOE’s and ask their help/guidance. Check if there’s a local tech school or university that offers Six Sigma courses. If you cannot find anyone local, then you should read about DOE’s. Come back if you need suggestions on what to read.
A few questions:
What kind of imagery? – ground level, overhead(airborne, recon satellite)
If overhead what altitude, visibility index, etc.
If ground – what distance.
Is the imagery raw input or has it been cleaned up (sharpened, blur removal, etc.)
Is the imagery from the same recording instrument?
What kind of decision are you trying to make with respect to final door position? A simple yes it is there no it isn’t or are you actually trying to locate the door coordinates? If it is a yes/no then you will be in the realm of logistic regression and your coefficients will express the odds of the presence of a door as a function of the presence or absence of a particular feature. If it is actual coordinates then you will have to take multiple image samples along the fuselage and test them for door presence.
As for type of design you will probably want to first run a saturated design but that isn’t going to be the main issue – the big issue is that of finding a combination of door
types that will fill the design matrix.
For example: let’s choose 3 parameters – Door window, door handle and door frame. The simplest design would be door window, door handle and door frame all expressed as a simple yes it is there no it is not. This means you would have to have a minimum of 4 specific window/handle/frame constructions
1. window absent, handle absent, frame present
2. window present, handle absent, frame absent
3. window absent, handle present, frame absent
4. window present, handle present, frame present
and within each of those builds you would need to have some min/max window size, handle size, and frame thickness as part of the definition of their presence in the image. Given what you are trying to do you would also want to include at least one more combination – that with no door present.
This is one of those cases where a black box regression method might be superior to a DOE analyzed using regular regression methods.
Some time ago I worked on a similar problem which involved the presence or absence of a given amount of fuel in an aircraft wing tank. I built a basic design for the 3D positions of the tank, tank configuration complexity, etc. We then put varying amounts of fluid in the tanks and used the measured outputs from repeated runs of the design to train a neural net. Once we had the net trained we tested it on random tank constructions in random orientations with a known quantity of fluid in the tank. It worked very well.