PlateAI: 3¢ Washington

AI-assisted plate position identification for the 1851–57 3¢ Washington

The 3¢ Washington stamp of 1851–1857 is one of the most intensively studied stamps in American philately. Carroll Chase was the first to plate it systematically, and his landmark work established the foundation for everything that followed. Over the decades, a small and dedicated community of specialists — including Wilbur F. Amonette, Richard C. Celler, and others — extended, corrected, and deepened that foundation, each building on the work of those who came before. Today that tradition continues through J. Bryan O’Doherty’s stampplating.com, which brought the accumulated knowledge of the field into the digital age.

Printed from 13 hand-engraved plates, each containing 200 individual positions, every one of the approximately 2,600 positions has a unique combination of characteristics resulting from the engraving process itself — making each position identifiable from the stamp alone.

PlateAI brings machine learning to this problem. Given a scan of a single stamp, the system analyzes specific diagnostic regions — corners, framelines, and label areas — and compares the visual signature against reference images from all 2,600 positions across 13 plates. It returns a ranked list of candidates with a confidence assessment, and tools to compare your stamp directly against vetted reference images side by side.

Under the hood, PlateAI uses a multi-branch convolutional neural network (CNN) trained with triplet loss — a technique that teaches the model to recognize visually similar positions as similar and dissimilar positions as different, rather than memorizing fixed labels. Each branch of the network specializes in a different region of the stamp, and their outputs are combined into a single similarity score used to rank candidates. Trained on approximately 20,000 scans from four curated collections, the model achieves 95% Top-1 accuracy and 99% Top-5 accuracy on clean four-margin stamps. On three-margin stamps accuracy drops by approximately 7 percentage points; on two-margin stamps, by roughly 18 percentage points.

PlateAI works on Scott #10, 10A, 11, 11A, 25, and 25A — Types I and II only. Type III and IV stamps (#26, 26A) were printed from different plates and are not supported.

Whether you are new to plating, an experienced specialist, or simply curious about what machine learning can do with a 170-year-old challenge — PlateAI is designed to be a useful tool and a starting point for exploration, not a final word. Visual comparison and philatelic judgment remain essential.

Credits

PlateAI was conceived, designed, and built by David Fussichen. It would not exist without the philatelic foundation laid by Carroll Chase, Wilbur F. Amonette, Richard C. Celler, and the specialists who followed — nor without the encouragement, expertise, and generosity of Robert J. Lampert and J. Bryan O’Doherty, who represent the living continuation of that tradition.

Release Notes

← Back

Release Notes

Interface and model version history

Instructions

1. Load Stamp Scan

Click the button to load a scan or photo of a single “patient” Type I or Type II 1851–57 3¢ Washington stamp. PlateAI handles both imperforate (#10, 10A, 11, 11A) and perforated (#25, 25A) stamps equally well. It was not trained on Type III or IV stamps (#26, 26A) and will return meaningless results if you attempt to plate one.

For best results, use a 1200 dpi flatbed scan. Higher resolutions are accepted but provide no accuracy benefit. Lower resolutions and cell phone photos are accepted, but accuracy degrades with image quality.

2. Align and Crop

PlateAI aligns all stamp images to a common orientation before analysis. It does this automatically by detecting the centers of the four rosettes. Automatic detection succeeds on roughly 75% of stamp images. When it works, you’ll see colored crosshairs overlaying the four rosette centers and a yellow box connecting them. If the alignment looks correct, click “2 — Align & Crop.” An aligned and cropped version of your stamp will appear in the left panel; click it to view full size.

Automatic detection fails most often when cancels obscure two or more rosette centers. When that happens, one or more crosshairs will be misplaced and the yellow connecting box will appear skewed. Click “Mark Rosettes Manually” and then click the four rosette centers in clockwise order: Top Left → Top Right → Bottom Right → Bottom Left. Once all four are marked, click “2 — Align & Crop.”

3. Plate the Stamp

Once aligned and cropped, your stamp is ready to plate. No attribute selection is needed on a first attempt — just hit the button. If PlateAI doesn’t find the correct position on the first try, you can narrow the field by selecting known stamp attributes in the left panel.

4. Interpreting Results

PlateAI returns a ranked list of the top 20 best-matching positions for your stamp, showing the plate, relief, type, and Scott number for each candidate, along with a similarity score between 0 and 1. This score is the cosine similarity between your stamp’s feature embedding and the reference embeddings at each position — higher is a stronger match.

The results header shows the top predicted position and Scott number, each with a confidence level: CERTAIN, HIGH, MODERATE, or LOW. These levels are based not only on the raw similarity score but also on the gap between the #1 and #2 candidates — a small gap means two positions are nearly indistinguishable to the model, which warrants caution. Poorer quality stamps and stamps with missing margins will tend to produce lower confidence results.

Accuracy predictions by confidence level:

• CERTAIN — 100% accurate
• HIGH — 100% accurate
• MODERATE — 96% accurate
• LOW — 75% accurate

CERTAIN and HIGH both achieved 100% accuracy in testing, but CERTAIN carries a significantly tighter statistical threshold — it requires both a high absolute score and a large gap over the #2 candidate. Think of HIGH as “the model is confident,” and CERTAIN as “the model is very confident and the next-best candidate isn’t even close.” For LOW confidence results, treat the top candidate as a starting point for visual comparison rather than a definitive answer.

The Scott number prediction considers all 20 top candidates, not just the top result, and is typically highly confident even when position confidence is MODERATE or LOW.

To copy results to the clipboard, click Copy. To log results across multiple stamps in a session, click Log — the current stamp is logged immediately. Download your session log as a .csv file by clicking CSV.

5. Compare

Once results are returned, click Compare on any candidate to open a visual comparison window. Your patient stamp is shown in color (if your scan was in color) alongside grayscale reference images from four vetted collections: Chase, Amonette Historic, Celler, and Lund-Amonette-O’Doherty (LAO). All images — patient and references — are cropped and aligned identically, making direct comparison meaningful.

Three tabs offer different comparison views:

• Corners — the four corner regions, where many position-distinguishing features appear
• L/R Framelines — left and right frameline strips, shown as 10% compression scans alongside full uncompressed strips
• Top/Bottom Framelines — top and bottom frameline strips, same format

Click any image to zoom to 200%. Use the arrows at the bottom to move your patient stamp next to whichever reference you want to compare against. Use the arrows at the top to cycle through the top 20 candidates. Full-size zoomable views of each reference stamp are available from the header of the compare window. Known position attributes (recuts, guide dots, anomalies) are shown for each candidate.

6. Refine

Most stamps will plate correctly on the first attempt without any attribute selection. If the correct position doesn’t appear at the top of the list, you can narrow the field by specifying what you know about your stamp. Basic attributes — color (Orange Brown or Not Orange Brown), Type (I or II), Relief (A, B, or C), and Plate — appear in the left panel. For finer filtering, click Advanced to specify recuts, guide dot locations, plate anomalies, and margin position.

Note that the model has no built-in ability to predict relief. Selecting the correct relief will typically have the largest impact on narrowing results when the initial confidence is low.

7. Position Explorer

Position Explorer is a reference tool, entirely independent of PlateAI’s AI identification capability. Use it to browse all 2,600 plate positions across 13 plates and compare reference stamps from the four vetted collections — with or without a patient stamp loaded.

Select a plate from the dropdown, then click any position to open the Compare window for that position. Use the attribute filters in the left panel (basic or advanced) to narrow which positions are shown as eligible — ineligible positions are grayed out and not clickable. The match counts at the top of the explorer show how many positions pass your current filters, both for the selected plate and across all plates.

If a patient stamp is loaded and aligned, it will appear alongside the reference stamps in the Compare window. You can load a stamp at any time without plating it — Position Explorer does not require the stamp to be identified first.