Stamp Image Recognition

An interesting app would be to take those stamp identifying books and make the information in it to be used for an app that helps to identify the country of the stamp.

Don's right...There have been other attempts recognition type software...for this recognition type software...It has been a long and pain full road with this endeavour..

filkata..What percent accuracy has this app produced...?

Robert

There have been recent advances relating to compression of digital matrices that make running large convolutional neural networks on smartphones possible. I have done patent work in this area. Obviously a cloud server-based solution would be better. My opinion is that a machine learning, neural network implementation would be essential to even basic stamp image classification (put a stamp image in, get a catalog number out). Shade recognition also should be possible with a large enough training set; creating and maintaining the training set is a non-trivial problem here. I don't view watermark recognition as essential. It is enough to quickly get to the right catalog number from which the collector can narrow down the variety.

With a large enough training set that includes used stamps with cancels, the "cancel problem" also should be manageable unless a black cancel completely obliterates the stamp. There will always be corner cases that fail but a high success rate seems likely, again subject to creating a very large training set.

Chris

While I love technology, I am a bit dubious. What Chris describes is certainly technically possible but I do not think the issue is the technology. As I am sure Chris knows, patents and commercialization are too very different things. The real challenges are commercial feasibility and implementation costs. This is not something that a single developer whips up in their spare time, a significant investment in NRE and infrastructure will be required. Add to this that philately has typically been 10-15 behind the technology curve and I doubt that we will see anything like what Chris describes being applied to philately for a long time.
Don

Don,
I agree, but we can dream can't we? Perhaps an existing CNN system simply needs the right training set and UI.
Chris

Absolutely, I love to dream. A lot of driverless cars will be implemented on city streets over the next 5 year and this was a dream for many of us not long ago.

Lately I have been dreaming about paper (as we know it) going away and being replaced with 'smart' paper. A Mylar type disposable paper which can be written on, drawn on, display videos, listened to AND is fully connected to networks. I already have done some research work in the area of human cognition (we have a patent in place with the help of University of Connecticut and a NIH grant). This evidence based research proves that we learn and communicate best when more than one processing system (visual, auditory, touch) is used. 'Smart paper' does not yet exist but this will be a real game changer!
Don

@Robert

As I mentioned I am yet to hook up exact statistics, but it does get it right for me 80% of the time, greatly speeding up the process of tracking my collection online.

@Chris
That would be amazing, though the training set would be pretty hard to come up with.

@Don
I have applied an existing modern image matching algorithm to a stamp database, combined with a low-cost server, and the result is quite good, even if it is not 100% correct all the time. But indeed I do not imagine someone throwing huge resources at this because the potential user base is not that big as say Facebook's.

Fikata mentioned the database only contains stamp images from 1957-present. I may be biased, but I would think older stamps more often would be difficult to identify vs. modern ones. The vast majority of modern stamps have English inscriptions, except for GB. That said, I see the potential value in this endeavor if it helps to speed the process of identification and allows you to add the stamp to a have- or wants-list.

I agree with you that older ones are probably harder to identify. I am looking to expand my program to be able to recognize all the stamps (for which I have images), it is simply a matter of server resources and therefore cost. As my main collection is on the topic of space I have chosen to start from 1957 onwards :)
If I do get a decent userbase that generates some ad revenue I will definitely be looking at this upgrade.

I can now share with you the app I have been working on - you can get it here for Android devices: https://play.google.com/store/apps/...ampdetective

Recent changes include user cropping of the image, which helps with the accuracy of the recognition. Also, it can now handle stamps from all years (currently over 630,000 are indexed.)

I am eager to hear your feedback!