Find a Complete Image from a Crop or Small Detail

Use Local · Keypoints when you have only part of a book cover, illustration, product, packaging pattern or design screenshot and need the complete image in your library.

The workflow is: prepare keypoint data → upload a reference → crop to the detail → select Local · Keypoints → set scope and threshold → search → inspect the complete candidate image.

This example was run in the released 3.13.31 client with three different landscape images parsed for Local · Keypoints. A 460×220 crop of rocks, water and trees from a 900×600 stream photograph was searched across the database with minimum similarity 0 and a limit of 200. The complete source, “风景-28-备份”, ranks first at 54.77%; the other two candidates show 9.30% and 1.01%. This small example demonstrates the workflow, not an accuracy estimate for other libraries.

Useful scenarios

• Books and publishing: a cover corner, spine, illustration or page fragment.
• Art and image libraries: brushwork, a decorative element or a distinctive pattern.
• Design and packaging: a cropped client screenshot, layout detail or printed graphic.
• Products and fabrics: a recognizable edge, structure or surface pattern.

1. Prepare local-keypoint data

When adding folders, enable Image Search and select Local · Keypoints among the strategies, then choose the source folder and click Start Adding.

The screenshot shows the feature choices before a folder is selected, so Start Adding is disabled. Choose your own source folder before starting.

For existing images without keypoints, enable Local · Keypoints under Settings → Add Files while retaining other needed parsing options, then right-click a few files and choose Reparse File. Keypoint repair only aligns already saved features with the index; it does not first-extract keypoints. Back up the database before reprocessing, which may affect existing collection associations and other image records. Confirm task success, open database details to refresh keypoint statistics, and reopen Image Search before expanding the scope.

Keypoint extraction can take substantially longer than ordinary global parsing. Try one project folder first and verify the results before processing a large library.

2. Open Image Search

Select Image in the workspace toolbar, then choose a file, drag an image into the dialog or paste a screenshot. Select Local · Keypoints. If it is unavailable, check the library's parsing data first.

3. Use an actual crop

If the reference is already a detail, use it directly. Otherwise choose Crop, select only the distinctive area and confirm. The reference shown back in the search dialog should now contain that crop, not the complete original screenshot.

Keep recognizable lines, edges, patterns or product structure. Exclude chat bubbles, borders, unrelated text and other products. The smallest possible crop is not always best: a flat color, blurred edge or repetitive texture may provide too little distinguishing information.

4. Set scope, threshold and result limit

Confirm Local · Keypoints, choose the minimum image similarity and maximum result count, then start within the database or a relevant project folder.

For an initial check, try 100–200 results. If too few appear, lower the threshold or widen the scope; if too many unrelated candidates appear, narrow the scope, raise the threshold or improve the crop. Increasing the limit cannot create data that was never parsed.

5. Adjust search level if needed

In released version 3.13.31, open Settings → Local · Keypoints. The Search section contains Search Level. The Search settings page also links to this panel.

Level 1 is faster with fewer candidates; level 10 searches more thoroughly but takes longer. Start with the default, then increase gradually if relevant matches are missing and parsing is complete.

Search Level affects later searches without re-extracting keypoints. The extraction limit and distribution are separate settings: they require a restart and reprocessing of existing images.

6. Verify the complete image

Open the first candidates and compare the crop with their complete image. Repeated fabric, wood grain, packaging families and similar patterns can produce plausible but incorrect matches.

Check that the detail occurs in the candidate, its direction and structure agree, and the file has the dimensions, quality and path needed for delivery. Similarity is a ranking clue rather than proof of identity or suitability.

7. Choose another strategy when appropriate

Use Global for a complete reference and overall appearance, Global · Subject when the subject matters more than the background, and Global · Color when the palette is the main requirement. Flat or heavily blurred crops are often poor keypoint references.

Troubleshooting and further tuning

If the strategy is unavailable, finish parsing its data. If the original is absent, check that the right library and folder are selected. Try a clearer crop with more distinctive detail before increasing every search setting.

In 3.13.31, Local · Keypoints results can display scores above 100%. These values are not accuracy percentages or proof of an exact match. Open the candidate and verify the detail, dimensions and path.

For scans or text-heavy pages whose detected keypoints cluster around lettering, see “Tune Local Search with Keypoint Previews”. Preview the distribution, apply changes only after comparison, and reprocess the relevant existing images before judging the effect on search.