Current Development Status of Image Recognition SW

Have you looked into OpenCV. Its fairly easy to make a wrapper for calls from LV. My next project is to update some old LV wrappers to the latest OpenCV.

The original code is from Intel.

Cheers

Some Haar training stuff...

Cheers

Down to the *species*?

The gross morphological differences between phenotypes within a given species are almost always comparable to the gross differences between species. Or in other words, just because it's bushy on top and woody on bottom doesn't tell you what species it is. Let alone other forms, like some species (more than others) can be trained into, or will naturally grow in, a bushy or even bonzai shape. To say nothing of arborism, heh.

Tim

-- Seven Transistor Labs Electrical Engineering Consultation Website:

Just for fun, looking into a DIY project, but which could spin off to a commercial product, I am looking into image recognition.

For the application I would need the unit to recognize an object. The reference object would be stored in a database with a picture from different angles, and the SW would compare an observed picture to the reference images and product a percentage match figure.

The image would not be a simple figure, but could be of different sizes and colors, but with the same fundamental shape. The image could for example be a tree, and the system would need to determine what kind of species it was and report the match figure.

So, knowing little of current developments, it this feasible or would that be 10 years into the future?

NI Labwindows has image recognition SW but I thing it will only cover basic shapes

Google has some advanced function in Google Photos:

day and that would then be a showstopper

Any insights?

Cheers

Klaus

This task is simple. It could take a lot of computing though.

Do you need to match entire object against database or do you have to split object into components to classify it?

It depends. Basically this is a question of how much computation you can afford and if there could be shortcuts to reduce computation.

Very basic and trivial stuff.

What do you really need?

Vladimir Vassilevsky DSP and Mixed Signal Designs

There are aps available right now for the blind. Use your android, or cellphone, to look at an object and it will tell you what the object is.

On a sunny day (Sat, 21 Dec 2013 14:59:26 -0800 (PST)) it happened Klaus Kragelund wrote in :

This seems fun. but have not tried it, its about face recognition, but there is a lot of open source stuff on the web:

It is quite difficult and requires a lot of computing power what you want to do. And if you are going to use open source libraries, please mind the license.

a dude in a callcentre identifies it and tells you (and NSA etc) what it is.

I don't think it classifies as simple, certainly limited material is available and what I have found sofar indicates that the success ratio of the recognition engine is not that impressive

Could be both, but I guess it would be better to have the entire object. If the algoritm splits the task up in smaller images is another matter.

It is for a robot. So I would aquire an image, send it to the server and let the server do data crunching and return with tags describing what is present in the picture. I do not think the function can be embedded into a simple embedded platform.

Well, I need a robot to be able to distinguish between different objects, that's the basic function

Cheers

Klaus

Image recognition is no magic. It is about correlation for all different scales and view and rotation angles. Albeight being very simple, that could take enormous amount of number crunching. The trick is finding shortcuts that would allow selecting areas of interest in the image; to reduce computation. If you could narrow down the search by locking on bright dots, lines, trivial geometric shapes, characteristic patterns or any clearly distinctive simple features, that helps.

There are two modes of image recognition: initial recognition and continuous tracking. Initial recognition works without prior knowledge, what is in the image. Tracking knows what is in the image so it only has to recognize changes as new object enter the picture. Of course, tracking is simpler task.

It depends.

How far, how many, how distinctive are the objects.

Vladimir Vassilevsky DSP and Mixed Signal Designs