For the grading job, needing a check of a continuous or almost continuous number, meaning the level of wear, a separate Convolutional Neural Network is used, this network being different from the one doing the main type identification, having been trained not for finding the coin type but for finding its specific condition and to check value of coins.
Task Types
First, as a Regression Model, the network gives a number score from 1 to 70, for example, 63.5, this score matching the standardized Sheldon grading scale, this task needing very careful settings and a large amount of training photos with exact number labels.
As a Multi-class Classification Model, the network puts the coin into one of the clear condition groups, for example, Fine, Very Fine, or Mint State.
Training these models means making a special set of data, with the same coin shown in many different pictures, these pictures marked exactly by expert coin collectors, showing all possible conditions from the MS-70 to the heavily damaged like Poor-1.
Extracting Wear Features
Checking Relief Graduation, the network looks at how smoothly the flat coin field changes to the highest points of the design, such as hair, feathers, or small letters, these changes becoming very flat on worn coins.
Checking Micro-element Clarity, the CNN checks how present and sharp the very small details are, these details being missing on worn coins, for example, thin lines in a beard or small decoration lines on the border.
Moreover, analyzing Field Texture, the program looks for the presence or lack of mint luster marks, this luster being the original surface texture made during the striking process.
Analysis of Key Zones and Defects
The process of checking condition always focuses on the coin areas most easily hurt by rubbing, these areas' condition being the main sign of the coin's grade.
Maximum Relief Zones
First, checking Detail Loss, the algorithm compares small image parts of these zones with perfect example patches, calculating the percentage of the original detail that is gone.
Second, checking Wear Depth, the program uses the brightness change data within the coin's design height to guess how flat the high points are, this flatness being a direct sign of physical rubbing.
Identifying Mechanical Damage
A separate thing lowering the grade, no matter the coin's general wear, is mechanical damage, having three types of checks.
The network looks for line defects with sharp brightness changes, these defects not matching the coin’s original design, being recorded as damage that happened after the coin was made.
It finds small local dents or holes, these damages often being round or cornered. Third, finding Cleaning Marks, parallel or circle-shaped small scratches are found, showing an attempt at mechanical cleaning, this action greatly lowering the collector price.
Patina and Surface
Checking Patina Evenness, the program looks at the color spread on the coin's surface, with spotty or too-bright patina possibly meaning strong chemical damage or bad storage, this bad condition having a negative effect on the grade.
Finding Mint Luster, the rest of the original metal shine is very important for high-grade coins, with the network looking for specific bright spots and texture, these signs only showing on coins that were never used in buying or selling.
Assigning the Numerical Grade
After getting all the numbers for wear, damage, and surface condition, this data is put together for setting the final coin grade.
Data Combining and Feature Weighting
The number scores, including the percentage of lost design height, the number of scratches, and the evenness of the patina, are given to the final part of the model, with different features having different "weight" depending on the coin type and its history, for example, wear being important for old American coins, while a lack of damage being key for new ones.
Calibration to the Sheldon Scale
The final result is set using the standard 70-point Sheldon scale, this scale being used in most grading systems.
This scale has clear levels:
MS, scoring 60–70, meaning the coin was never used, with the AI looking for full design height and mint shine, the differences between MS-63 and MS-65 being set by the number and visibility of very small production errors.
AU, scoring 50, 53, 55, 58, means the coin was almost never used, having only small signs of rubbing on the highest points, with the AI finding a little bit of flatness.
VF, scoring 20, 25, 30, 35, shows much wear, but the main parts of the writing and design are still clear.
G and P, scoring 1–12, means the coin is very worn, with most small details and writing being lost, but the coin type still being found, the number result given by the model being rounded to the nearest standard grade.
Determining Market Value
Finding the collector price is the last step, depending on three exact numbers: the coin’s unique ID, its grade, and the current market data.
Linking to the Numismatic Database
For data sources, the app collects price facts from open auction lists, sales books from big sellers, and coin price lists.
For price updates, the price numbers show the current market changes, with the price the app shows for an MS-65 coin being the middle or average price of the most recent good sales of that coin in that condition.
Factors Affecting the Final Price
The final price report looks at several things besides the basic grade, with the first factor being Mintage, meaning the number of coins made, where a smaller number of coins made leads to a higher price, with the app keeping this history data.
The second is Rarity of Variety, meaning if the coin is found to be a rare die type, the price is moved up, this finding being done by the OCR or detection module.
The third is Additional Marks, where color names are important for some coins, like "Red" for copper coins, with the DL model being able to find and group the patina color, this grouping being used to make the price more exact.
Ensuring and Improving Grading Accuracy
High accuracy in checking coin condition is a very hard technical job, needing constant training and setting of the system.
For making sure the grading model works exactly right, no matter the conditions when the user takes the coin photo, developers use data augmentation, where the original perfect coin images are changed on purpose, for example, by turning, darkening, or lightening them, or adding fake noise or shadows, with the model learning to see the real wear while ignoring these fake changes.
As for Focusing on Texture, where the model is specially trained to check local texture patterns, not the overall brightness, training makes it stable against uneven light.