Evolution and Technical Analysis of Dewu Photo Search

Introduction: Users often encounter difficulty identifying products from images; Dewu offers a photo search feature that matches uploaded images to similar items.

What is photo search: It uses semantic understanding of images to retrieve matching products; similar services include Google Image Search, Baidu Image Search, Alibaba's Paillitao, JD, Sogou, 360.

Dewu Photo Search Overview: Initially a small feature with limited categories; the V1 version (Aliyun) relied on a single image library and category‑specific detectors, leading to issues such as non‑shoe items being returned for shoe queries.

Problems Identified: Inaccurate target detection caused wrong category results and poor positioning; analysis showed multiple detectors and weighted voting contributed to errors.

V2 – Optimized Aliyun Search: Expanded categories (e.g., clothing), added a custom detector to pre‑filter the product class, and improved accuracy by ~5% overall, with notable gains for shoes (+41%) and clothing (+11%). Added self‑developed detector increased latency to ~600 ms, exposing stability concerns under high load.

V3 – Self‑Developed Photo Search: Designed a full pipeline—target detection, feature extraction (EfficientNet backbone), embedding to 128‑dim vectors, and similarity search. Employed letter‑box resizing, data augmentation, triplet loss, and cross‑entropy loss. Achieved >100% improvement in top‑1 shoe accuracy compared with the Aliyun version.

Performance and Efficiency: Self‑developed service removes base64 transfer, reduces processing steps, and supports higher QPS, but still faces challenges in feature computation and scalability.

Conclusion: Photo search integrates computer‑vision techniques and is becoming a key user‑driven product search method; continued optimization of detection, feature quality, and system stability is essential for future growth.