Client‑Side Image Similarity Computation: Methods, Experiments, and Findings

In e‑commerce scenarios, duplicate product listings waste storage and compute resources. Performing image similarity detection on the client before uploading can filter repeated images and videos, saving bandwidth and cloud storage.

This work evaluates three categories of image‑similarity techniques—perceptual hash algorithms, local invariant‑feature matching, and convolutional‑neural‑network (CNN) based methods. By comparing computational complexity and retrieval efficiency, the study selects Hessian‑Affine feature detection combined with SIFT descriptors as the preferred on‑device solution.

Traditional hash methods (average hash, difference hash, perceptual hash) generate compact binary fingerprints; similarity is measured by Hamming distance. Experiments show that while fast, these hashes provide insufficient accuracy for practical duplicate‑detection tasks.

CNN‑based end‑to‑end similarity (e.g., using a VGG16 backbone) extracts deep features and feeds them to a fully‑connected matcher, achieving high accuracy but incurring prohibitive compute cost for large‑scale video retrieval. An alternative is to use intermediate feature maps (e.g., VGG16 block5_pool) and compute Euclidean distances, offering a trade‑off between accuracy and speed.

Local invariant‑feature approaches such as SIFT and Hessian‑Affine are examined. SIFT demonstrates strong robustness to cropping, subtitle overlay, and brightness changes, but struggles with rotation. Introducing a minimum key‑point threshold (e.g., 30 points) markedly improves precision by discarding low‑confidence queries.

Video‑frame experiments (10 k videos, 10 frames each) reveal that Hessian‑Affine + SIFT raises both recall and precision while reducing the exclusion rate to about 4 %, outperforming plain SIFT.

Conclusion: For on‑device image similarity, Hessian‑Affine detection paired with SIFT descriptors offers the best balance of computational cost, retrieval efficiency, and robustness. Future work may integrate local features with deep semantic features to further enhance performance.