Message:

<p>[QUOTE=&quot;RonSanderson, post: 7332199, member: 77413&quot;]At the current level of technology, all AI means is that you have a generic system that can review large sets of data and extract correlations.</p><p><br /></p><p>Let’s say we had a single Lincoln cent, and that everyone agreed it was the most perfect example of an MS65 that all grading services certified. First off, you would need to have a controlled photographic setup, so the lighting was consistent and did not obscure any details. Then you would want to photograph the coin from multiple angles, probably with two or three lights to help bring out the contours as bright highlights and darker shadows.</p><p><br /></p><p>Next you have a couple of approaches.</p><p><br /></p><p>One would be to feed all the images into a program that aligns and matches them to create a mathematical model of the surface. (Homographic tomograpy). Then you could start adding more coins, photos, and data to build and idealized model of the coin. Algorithms exist to compute the variance of an image from the ideal model. This could be tuned so that the more degraded the markers in the image, the lower the grade.</p><p><br /></p><p>The second approach would be to “train” an AI package to detect similarities and differences between coins. You would accumulate as many sets of data as possible by, again, photographing as many coins as possible under precise conditions. Feed in each set of data, along with the grade, denomination, and possibly the date/mint mark, and let the AI software identify correlations between the data sets and the grade. It may well find that some poorly produced years may have special exceptions to the overall deduced rules.</p><p><br /></p><p>This is just like the Captcha images you often see - “Which squares contain a traffic signal” - or hills, or road signs. Your responses actually help train recognition software, from what I read.</p><p><br /></p><p>After building the dataset, you can start feeding in new images and letting the software respond with the grade. After having done this enough times, and then feeding back into the system the actual grade, you should start converging on a reliable process.</p><p><br /></p><p>Of note here is that more data makes better results. And better, widely agreed upon grades are essential to building a reliable data store with dependable rules.</p><p><br /></p><p>Would this need to be done separately for each series? Is there enough available data to train the system? Can the system pick up the essence of MS65 if you feed in a mixture of coin denominations? Are all gold coins the same? Copper Silver?</p><p><br /></p><p>Those questions would come out of the attempt to load data, but whether they affect the reliability would have to be determined as the project went along.[/QUOTE]</p><p><br /></p>

Your name or email address:

Do you already have an account?

• No, create an account now.
• Yes, my password is:
• Forgot your password?

Stay logged in