Implementation and Optimization of a Political Face Recognition Service

The article introduces a political face recognition service that filters images containing political figures at an industrial scale, using a three‑stage pipeline of face detection, face alignment, and face recognition.

Business growth has led to billions of images needing automated political‑sensitive screening; manual review is infeasible, prompting the development of a self‑service system with adjustable thresholds and a sensitive‑person database.

The technical solution consists of detecting faces, aligning them, extracting 512‑dimensional features, and comparing them against a pre‑loaded political‑person feature library using cosine similarity. Various detection algorithms (MTCNN, FaceBoxes, SSH, FasterRCNN) were benchmarked; FaceBoxes was selected for its balance of speed and accuracy.

For alignment, a cascade of three CNNs (PNet, RNet, ONet) is used, and the Wing loss function (with parameters w=10, a=2) is adopted to improve landmark regression robustness.

Recognition relies on the ArcFace loss with a ResNet‑100 backbone, chosen after comparing several loss functions and networks on the LFW benchmark, where ArcFace achieved the highest accuracy (99.83%).

Optimization efforts include enhancing the FaceBoxes network by adding an Inception module and replacing pooling layers with strided convolutions, raising recall from 82% to 98.3% while reducing latency from 35 ms to 30 ms. Recognition was further improved by cleaning public datasets (MS‑Celeb‑1M, VGGFace2, CelebA, etc.) and augmenting them with a proprietary 58Face dataset; MobileFaceNet was adopted as the backbone, yielding 99.83% LFW accuracy and 99.98% on the internal test set, with inference speed of 30 FPS on GPU.

Deployment results show the service reaching a recall of 59.89% after four iterations, supporting over 140 business scenarios and handling peak daily request volumes of 350 million calls.

The conclusion highlights remaining challenges such as scaling the political‑person database, handling diverse lighting, pose, and occlusion conditions, and plans to expand training data, explore new alignment methods, and build a comprehensive face‑quality assessment framework.