image recognition

[ntahlah]

Hello! I'm a total newbie here and currently I'm planning to do a plant defect detection thru image recognition for my project (basically I just want to detect whether the plant got disease or not). I just wondering does face recognition concept can be apply in this project?

Thank You!

[forenbenutzer]

This might work, but as far as I remember, you need tons (several thousand) sample images to "teach" a classifier. Or you make use of several (weihgted) classifiers.

This also might be a first start:
http://en.wikipedia.org/wiki/Principal_ ... t_analysis

[PiGraham]

What does a diseased plant look like?
Plants can be very difficult to classify because they are so variable. If the leaves change colour you may be able to detect that quite simply. Maybe density of foliage would be an indicator. Maybe water uptake or CO2 absorption would help.
Different plants and different diseases in different conditions may all have different signs of disease.

Most likely it will be very, very difficult to get reliable results.

Face recognition will not help.

If you are thinking of a general purpose plant disagnosys device that just works on most plants in any circumstances forget it.

This is not a newbie project.

[ntahlah]

What does a diseased plant look like?
Plants can be very difficult to classify because they are so variable. If the leaves change colour you may be able to detect that quite simply. Maybe density of foliage would be an indicator. Maybe water uptake or CO2 absorption would help.
Different plants and different diseases in different conditions may all have different signs of disease.

Most likely it will be very, very difficult to get reliable results.

Face recognition will not help.

If you are thinking of a general purpose plant disagnosys device that just works on most plants in any circumstances forget it.

This is not a newbie project.

First of all I would like to thank you for reply.. actually I just use a paddy plant for my subject I just want to detect brown spot and leaf blight on the leaf as far as what I have researched most of this disease tend to have spot on the leaf which I think excellent to use image recognition, so as long as it can detect the spot on the leaf I think it's ok

[ntahlah]

This might work, but as far as I remember, you need tons (several thousand) sample images to "teach" a classifier. Or you make use of several (weihgted) classifiers.

This also might be a first start:
http://en.wikipedia.org/wiki/Principal_ ... t_analysis

Thanks for the info!

[PiGraham]

First of all I would like to thank you for reply.. actually I just use a paddy plant for my subject I just want to detect brown spot and leaf blight on the leaf as far as what I have researched most of this disease tend to have spot on the leaf which I think excellent to use image recognition, so as long as it can detect the spot on the leaf I think it's ok

Is this to be a fully automatic system? How would the system be used?
Reliably finding leaves and detecting disease spots on them is far from easy.

[ntahlah]

First of all I would like to thank you for reply.. actually I just use a paddy plant for my subject I just want to detect brown spot and leaf blight on the leaf as far as what I have researched most of this disease tend to have spot on the leaf which I think excellent to use image recognition, so as long as it can detect the spot on the leaf I think it's ok

Is this to be a fully automatic system? How would the system be used?
Reliably finding leaves and detecting disease spots on them is far from easy.

Yes I'm planning to do a fully automated system, so basically this system will be attach in front of a moving tractor and then the tractor will swipe thru the paddy field while the system will scan all the paddy plant on the field.. but hey I just wondering since all the healthy leaf will be green coloured what if I just let the Neural network learn the the property of the healthy plant and the unhealthy one will be just consider as defected so I just need to use colour recognition right?

[PiGraham]

Yes I'm planning to do a fully automated system, so basically this system will be attach in front of a moving tractor and then the tractor will swipe thru the paddy field while the system will scan all the paddy plant on the field.. but hey I just wondering since all the healthy leaf will be green coloured what if I just let the Neural network learn the the property of the healthy plant and the unhealthy one will be just consider as defected so I just need to use colour recognition right?

That's a tough application and I really don't think the Pi is capable of that. Consider a single row of plants. How many leaves would the system have to identify and inspect every second?
What are the chances of presenting individual leaves one at a time face on to the camera?
What proportion of diseased leaves can the system miss and still be useful?
What proportion of healthy leaves can the system mistake as diseased and still be viable?
What happens when disease is detected?

IF the leaves are easy to see and the disease spots are very obvious, and you can tolerate a fairly high error rate, and it isn't fast
This sort of application is a specialist task. I think you would need expert help with this.

You will surely need multiple Pi's to have any chance of covering the width of a tractor. You can probably get better price/performance with a more powerful system.

Pi isn't suited to interfacing to cameras other than the Pi camera, and that isn't well suited to this application.

[ntahlah]

Yes I'm planning to do a fully automated system, so basically this system will be attach in front of a moving tractor and then the tractor will swipe thru the paddy field while the system will scan all the paddy plant on the field.. but hey I just wondering since all the healthy leaf will be green coloured what if I just let the Neural network learn the the property of the healthy plant and the unhealthy one will be just consider as defected so I just need to use colour recognition right?

That's a tough application and I really don't think the Pi is capable of that. Consider a single row of plants. How many leaves would the system have to identify and inspect every second?
What are the chances of presenting individual leaves one at a time face on to the camera?
What proportion of diseased leaves can the system miss and still be useful?
What proportion of healthy leaves can the system mistake as diseased and still be viable?
What happens when disease is detected?

IF the leaves are easy to see and the disease spots are very obvious, and you can tolerate a fairly high error rate, and it isn't fast
This sort of application is a specialist task. I think you would need expert help with this.

You will surely need multiple Pi's to have any chance of covering the width of a tractor. You can probably get better price/performance with a more powerful system.

Pi isn't suited to interfacing to cameras other than the Pi camera, and that isn't well suited to this application.

Hmm I guess you have a point there... i guesss i have to find another way around.. anyway I would like to thank you for your time to answer all my questions here..

Have a great day!

[PiGraham]

What you could do is to experiment off line with Pi and imaging libraries like opencv. Seeing something up on the bench and experimenting with presentation, lighting and processing is a good way to learn. You can scale up from there. If you have the time and the enthusiasm and you are not under pressure to get plant disease detection working then it can be a very interesting thing to work on.

I suggest you start with some simple projects first.

[you452]

Have you found another way as you said ... (I am also working on similar project)

[ntahlah]

Have you found another way as you said ... (I am also working on similar project)

Hi there! What I did is as what PiGraham suggested which is by doing it off line instead of doing it automatically as per what I planned before (note that during that time I was using Raspberry Pi 2, that thing doesn't have the processing power to do what I planned to do).

So what I did was I extracted the features of the leaf first by using Raspberry Pi capabilites and preprocess the feature as well. The result of the proeprocess image will then submitted into PC for classifying process (where I apply my machine learning algorithm). Did the job though... (technically)

But since technology has vastly improve, I suggest you explore on Deep learning to do the process as current Raspberry Pi (Raspberry Pi 3 Model B) can handle the load (I have tried it but on a different subject).

Hope this answer your question

[PiGraham]

https://svds.com/tensorflow-image-recog ... pberry-pi/

[MaxK1]

Does IR reveal more damage to a plant than visible light? If so, then maybe the pi camera without the IR filter might be better.