Deep Learning for Automated Module Detection in Taobao 99 Promotion Pages

Introduction: The 99 promotion on Taobao involves thousands of pages; testing is resource‑intensive. This work explores using machine learning to automate page module testing.

Process: Deep learning pipeline consists of model training and prediction. Supervised learning requires labeled samples and appropriate network selection.

样本生成：样本生成过程中使用的三种方法

模型选择：fpn + Cascade R-CNN

Prediction includes module recognition, element recognition, and position refinement to achieve pixel‑level accuracy.

模块识别：识别页面中模块类别和位置

元素识别：基于模块的识别区域，识别内部元素

位置修正：通过传统图像的方式修正内部元素位置

Model selection: To meet high IoU recall and precision (≥95%), a Cascade‑RCNN with Feature Pyramid Network (FPN) is employed, improving detection of varied module sizes.

Results on ~1000 validation images (COCO metrics) show AP up to 0.989 and AR up to 0.993, indicating ~98% recall and precision for module and element detection.

Future work: Extend the approach to other large‑scale events (e.g., Double‑11) and balance accuracy with inference latency.