Similar Image Collection Review

How to use Similar Image Collection Review

Review perceptual-fingerprint candidate pairs across a small local image collection, with exact Hamming distances and explicit false-match limits. Hash bounded image structure, not filenames or a semantic identity model. The JSON report lists threshold edges and connected groups without deleting files or declaring them exact duplicates.

1. Provide static images to compare, maximum 128-bit hamming distance.
2. Choose the supported mode and run the tool. Review its assumptions and any reported errors.
3. Copy the result or download a separate output; reset to clear the current form.

Good to know

• 2–16 local files, each at most 8 MiB, 1,000,000 pixels and 2,048 per side; 24 MiB and 4,000,000 source pixels combined. The original two-file example deliberately contains identical fixture bytes under different names. Static non-animated PNG, baseline JPEG or simple VP8/VP8L WebP only. Structural/CRC preflight precedes native decoding. APNG, progressive JPEG, extended/animated WebP, compressed PNG metadata, SVG/GIF and orientation changes that disagree with the header are refused.
• Authored dHash-style 128-bit descriptor: white-composite decoded alpha, weighted encoded-sRGB gray, exact area averaging to a 9 × 9 grid, then horizontal/vertical strictly increasing comparisons over the top-left 8 × 8 cells, row-major most-significant-bit first. This is not claimed bit-compatible with another library’s grayscale/resampler.
• The gray weights are 2126 R + 7152 G + 722 B over encoded channels, not a linear-light luminance or medical vision model. The area-weight integer arithmetic is exact within the 1-million-pixel cap. Sampling work is at most (source width + 8) × (source height + 8) cell intersections.
• An integer threshold 0–128 includes a candidate edge when XOR popcount is at most that value. All at most 120 pairs are considered; connected groups may contain endpoints that do not themselves pass the threshold. The report retains those direct edge witnesses instead of declaring every group mutually equal.
• A descriptor is not a cryptographic checksum, byte equality, ownership, identity, authenticity or reliable semantic similarity. Uniform but different-colored images can collide; crops, rotations, overlays and composition changes can also cause misses. No threshold proves that deleting a file is safe; compare the originals visually and use the exact duplicate-file tool for byte equality.
• Each descriptor runs in a prepared same-origin module Worker: up to 10 seconds to prepare, then a separate 2-second computation deadline per image. Native decoding waits at most 10 seconds and resources are released. Nothing is uploaded, opened by URL or automatically removed.
• Structured text output is limited to 500,000 characters. Reduce the supplied source when that separate output budget is exceeded.
• Inputs are processed on this device and are not automatically saved or uploaded.