The Unrealized Promise Of Multifrequency

[mh9162013]

4 hours ago, Skookum said:

I keep thinking the next level of MF is going to come from increases in processing power that allow application of AI algorithms within the machine.  What if a processor analyzes how a target responds to multiple frequencies or technologies at the same time and then it extrapolates that information to identify other targets at your location?  No doubt this already happens in R&D labs, but could there be fresh gains having the process occur in real-time, on the ground?  Perhaps, its another potentially cool application of artificial intelligence.

I've had similar thoughts.

For example why can't the metal detector be programmed for the user to tell the machine what was found? After the user does this enough, the detector will have "learned" how to identfy targets in the ground based on what it was detecting and what was ultimately dug up.

It can be limited to just American coins, for instance. Let's say a user (I'm gonna used Minelab's MIQ scale here) finds a target that rings up as a 25 and is 4 inches down. The user thinks it might be a copper penny. The user carefully digs it and notes how deep it was, its orientation in the soil (if that's possible to discern...it sometimes is when the coin is neatly stuck to the side or bottom of the plug) and what the object actually was.

So in my hypothetical, it turns out to be a 2004 Zincoln, not a copper penny. The detector will make a "mental note" of the find and several variables that corresponded with it (including the GPS coords of the find). If a user is willing to take the time to teach its detector to do this, the machine would be pretty darn good after just a dozen or so digs in the same general location.

The problem I see with this approach is threefold. First, it requires a lot of time and patience. Second, there's the problem of "garbage in, garbage out." If the user is wrong in the data it gives the machine (gives the wrong depth of the coin, for instance, when the user swings the coil higher off the ground than is recommended), then the detector can't properly learn. Third, you're basically teaching a machine to do what a neurotypical human being who's really into metal detecting is doing anyways. Basically, this hypothetical machine isn't just a metal detector, but a notetaker and pattern finder. But humans are already pretty good at seeing patterns. They're just lousy with memory and tend to be lazy in their habits. So it's too easy for them to incorrectly remember: the find, the settings used and the ground conditions at that time.

A simpler process would just be to use the machine, tell the machine to note a given signal, then after digging it up, simply telling the machine if it was a "good" or "bad" target. This boolean "process" could also teach the detector about what's a potential good find or not. For coin shooters (gven how standardized the desired targets tend to be), this could be pretty effective. I think in its ideal form, it would still give a tone and VDI like any other detector, but it would have a "confidence bar" to go along with it. It's this confidence bar that would be constantly altered as the machine learns from the user.

[☠ Cipher]

15 hours ago, mh9162013 said:

I've had similar thoughts.

For example why can't the metal detector be programmed for the user to tell the machine what was found? After the user does this enough, the detector will have "learned" how to identfy targets in the ground based on what it was detecting and what was ultimately dug up.

It can be limited to just American coins, for instance. Let's say a user (I'm gonna used Minelab's MIQ scale here) finds a target that rings up as a 25 and is 4 inches down. The user thinks it might be a copper penny. The user carefully digs it and notes how deep it was, its orientation in the soil (if that's possible to discern...it sometimes is when the coin is neatly stuck to the side or bottom of the plug) and what the object actually was.

So in my hypothetical, it turns out to be a 2004 Zincoln, not a copper penny. The detector will make a "mental note" of the find and several variables that corresponded with it (including the GPS coords of the find). If a user is willing to take the time to teach its detector to do this, the machine would be pretty darn good after just a dozen or so digs in the same general location.

The problem I see with this approach is threefold. First, it requires a lot of time and patience. Second, there's the problem of "garbage in, garbage out." If the user is wrong in the data it gives the machine (gives the wrong depth of the coin, for instance, when the user swings the coil higher off the ground than is recommended), then the detector can't properly learn. Third, you're basically teaching a machine to do what a neurotypical human being who's really into metal detecting is doing anyways. Basically, this hypothetical machine isn't just a metal detector, but a notetaker and pattern finder. But humans are already pretty good at seeing patterns. They're just lousy with memory and tend to be lazy in their habits. So it's too easy for them to incorrectly remember: the find, the settings used and the ground conditions at that time.

A simpler process would just be to use the machine, tell the machine to note a given signal, then after digging it up, simply telling the machine if it was a "good" or "bad" target. This boolean "process" could also teach the detector about what's a potential good find or not. For coin shooters (gven how standardized the desired targets tend to be), this could be pretty effective. I think in its ideal form, it would still give a tone and VDI like any other detector, but it would have a "confidence bar" to go along with it. It's this confidence bar that would be constantly altered as the machine learns from the user.

It is fun to wonder what would happen if you could incorporate artificial intelligence machine learning into these devices. I would imagine such things aren’t the strong suit of metal detector engineers/programmers and they might have to utilize outside consultation, but some level of this could certainly be done, and enabled to be switched on or off. 

[longbow62]

I'm speculating here. I think at the present time there would not be enough info gleaned about a target for a detector to learn much. A coin at 2" vs. 10" inches is even less info. If you were blinded folded and handed targets both coins and trash you could feel them to learn what each were. Now put on rubber gloves. It would be a little harder to tell. If you continually increased the thickness of the gloves you were wearing it would get harder and harder to tell what you were feeling. Eventually you could tell you were holding something, but you would have no idea what it was. Glove thickness equates to depth. Then you have bad soil which would equate to having to stick your hands down into sand to feel the targets.

From what I understand about how detectors work there is not that much info about targets to begin with just based on the electrical signal returned to the detector. I'm really surprised they are as advanced as the are and can tell you as much as they do about whats in the ground.    

[GB_Amateur]

At the risk of veering too far off the initial post (and title), one thing I've wondered about is if a detector could triangulate a target's location during the swing.  It would be a big help for me when coin detecting to know that something giving me a solid coin signal was actually so deep as to not be a coin -- would save me considerable digging time.  (Now I can figure it out before reaching the target, but it still takes a while before that realization occurs.)

My guess is that if this were possible it would have been done already.  But that attitude leads to zero innovation....

BTW (even more off-topic?), Garrett claimed their GTI 2500 could image targets.  I think they had (maybe still do) a special searchcoil with an extra loop of wire in it for the purpose.  But it never seemed to catch on, making me think the ability was overstated.  Must be a story (posts?) out there that I've missed....

[phrunt]

2 hours ago, GB_Amateur said:

BTW (even more off-topic?), Garrett claimed their GTI 2500 could image targets.  I think they had (maybe still do) a special searchcoil with an extra loop of wire in it for the purpose.  But it never seemed to catch on, making me think the ability was overstated.  Must be a story (posts?) out there that I've missed....

The Nokta Invenio goes some way to doing that, it's not perfect of course but it does image targets to an extent.

A friend has got one, and this is him using it demonstrating that

Maybe now they need an Inventio 2 Legend Edition with multi frequency to even further enhance its ability to do this, seeing Multi freakers have better target identification it should improve it over this early model.  If they do that he'd be annoyed he forked out for this one, it wasn't cheap.

[Jeff McClendon]

On 3/28/2022 at 10:32 AM, Steve Herschbach said:

So stop with the “multifrequency finds all targets across the entire range” nonsense. There is no multifrequency detector made that runs in one mode, and hits tiny gold nuggets, while also working perfectly well at the beach. That’s what we are being sold, and it’s simply not true.

I am going to assume that this quote is referring to Beach Sensitive. I am one of the idiots that said at least on paper, Beach Sensitive is like having a gold prospecting mode for a saltwater beach. Since I have never used a Deus 2 let alone Beach Sensitive at a gold prospecting site or Beach Sensitive at a saltwater beach on sub .1 gram nuggets............I have no clue except that I assume that the salt ground balance values for the frequencies used are going to seriously effect tiny gold nugget detection in the sub .1 gram range or any target that has a VDI near the salt ground balance values

[GB_Amateur]

45 minutes ago, phrunt said:

The Nokta Invenio goes some way to doing that, it's not perfect of course but it does image targets to an extent.

A friend has got one, and this is him using it demonstrating that....

...It wasn't cheap.

Good intro video for a more/less just out-of-the-box device.  It seems the target he demo'ed was an easy one (nice and large, good shape and orientation in the ground) and he mentions that smaller targets can be more difficult to image although he added that he found a (jewerly) ring which did display well.  Currently this is just over $5k here in the USA (from Serious Detecting).  I guess the price is why we haven't heard much about it.

Interesting that he said it learns if you feed it the right info.  I'm surprised because I would have expected it to require an unreasonable amount of training data to pull that off, but maybe it's limited to perfect targets, well away from trash and optimally oriented.  (Just my speculation -- I have no idea how 'smart' it really is.)  Once again Nokta/Makro dives into a challenging project and is able to release a functioning product in a reasonable amount of time.  The future looks bright for them and us.

[phrunt]

Yea, he finds it fun to use but doesn't seem to use it all that much ? So make of that what you will.

[Steve Herschbach]

1 hour ago, Jeff McClendon said:

I am going to assume that this quote is referring to Beach Sensitive. I am one of the idiots that said at least on paper, Beach Sensitive is like having a gold prospecting mode for a saltwater beach. Since I have never used a Deus 2 let alone Beach Sensitive at a gold prospecting site or Beach Sensitive at a saltwater beach on sub .1 gram nuggets............I have no clue except that I assume that the salt ground balance values for the frequencies used are going to seriously effect tiny gold nugget detection in the sub .1 gram range or any target that has a VDI near the salt ground balance values

Well, my whole tirade is more centered around this one idea - “This means you only need to cover the ground once and can be confident you’re not leaving ANY valuable treasure behind.”