Duplicate Photo Finder with Visual Similarity

Phone galleries silently accumulate duplicates: burst mode fires 20 shots to get one good one, WhatsApp forwards re-save the same image under new names, and "final_final_edit.jpg" sits next to the original. Byte-exact duplicate finders miss all of these — the files differ, the photos don't. PixelPrune works the way human eyes do: it shrinks each photo to 9×8 grayscale, compares each pixel with its right neighbour, and builds a 64-bit perceptual fingerprint (dHash). Two photos of the same scene produce fingerprints that differ in only a handful of bits even after cropping, re-compression or light retouching; two different scenes differ in roughly half. The Hamming distance between fingerprints becomes a similarity percentage, an adjustable threshold groups the library, and the compare view shows the pair side-by-side with the differing fingerprint bits outlined in amber. Deletion is deliberately conservative: the largest file per group is pre-marked "keep", the user ticks the rest, and everything goes to the recycle bin. The demo is a working web prototype of the desktop tool, running the real dHash pipeline on real photographs in the browser.

1. The user selects folders; the scanner enumerates image files (JPG, PNG, HEIC, WebP) — nothing is uploaded, all hashing is local.
2. Each photo is downscaled to 9×8 grayscale and each pixel is compared with its right neighbour to build a 64-bit dHash fingerprint.
3. Pairwise Hamming distances are computed; photos within the threshold's bit budget are clustered into duplicate groups.
4. Groups are ranked by similarity; the largest file in each group is marked KEEP and the reclaimable MB is totalled.
5. The user reviews groups, opens the compare view for any pair to see the similarity score and differing hash bits.
6. The user ticks unwanted copies; the delete bar shows the count and MB to be freed.
7. Confirmed deletions move files to the recycle bin; the scan summary updates with the new library state.

Tech stack:

• Python (desktop hashing engine + scanner)
• dHash perceptual hashing (64-bit fingerprints)
• Hamming-distance clustering
• Tkinter / PyQt desktop UI (built-to-order app)
• HTML5 · CSS3 · JavaScript (working web demo of the pipeline)
• Canvas-based hash computation in the demo
• Pillow (image loading and downscaling)
• Git version control

• [Perceptual hashing engine] 64-bit dHash fingerprints computed per photo — robust to re-save, resize and light retouching.
• [Visual similarity grouping] Photos clustered by Hamming distance with a similarity percentage per group (e.g. 97% similar).
• [Adjustable threshold slider] Lower values catch looser matches, higher values only near-identical shots — groups re-compute instantly.
• [Side-by-side compare] Any pair opens large with both 64-bit fingerprints shown and differing bits highlighted.
• [Smart keep suggestion] The largest file in each group is auto-marked KEEP, maximizing the space saved per deletion.
• [Reclaimable-space accounting] Live MB totals per group and overall, so the payoff of cleanup is visible before deleting.
• [Safe delete] Selected copies move to the recycle bin, never hard-erased; "keep at least one per group" is enforced.
• [Folder-scoped scanning] Scan Pictures, Downloads or a phone DCIM folder, with subfolders included automatically.

• Complete desktop application source code (Python)
• Perceptual hashing + clustering engine
• Scan results, compare and settings screens as designed
• Working web demo running dHash on real photos in-browser
• Sample 10-photo library (3 near-duplicate pairs + 4 unique)
• Project report PDF (problem, dHash theory, clustering, complexity, testing)
• PPT presentation for final review
• Viva Q&A preparation document (perceptual vs cryptographic hashing, Hamming distance, threshold trade-offs)

• Similarity is visual, not semantic — two different photos of the same sunset may group at low thresholds; the slider and compare view exist for exactly this.
• Heavy edits (major crops, filters changing most pixels) raise the Hamming distance and may escape the default threshold.
• HEIC support depends on platform codecs on the buyer's machine.
• The demo runs in-browser on a 10-photo sample; the desktop build scales to full libraries with an indexed hash cache.
• Face-aware "same person" grouping is not included — this is pixel-similarity dedup, not face recognition.