Edge Intelligence for Intelligent Video Cover Recommendation

This article presents the design and deployment of an edge‑based intelligent video‑cover recommendation system for the DeWu platform.

Background : Video covers significantly affect click‑through rates (CTR) for both creators and viewers. Automating cover selection reduces creator effort and improves CTR.

Edge Intelligence : Edge (client) intelligence performs data processing and inference on devices such as smartphones, offering low latency, bandwidth savings, privacy, and reliability. Challenges include limited compute resources, data consistency, and device management.

SDKs and Framework : The system uses the open‑source MNN SDK (supporting TensorFlow, Caffe, ONNX, etc.) to build a unified inference infrastructure that handles model download, caching, execution, and monitoring on both iOS and Android.

Overall Architecture : A shared client‑intelligence module provides model management, inference, and performance monitoring. The workflow extracts video frames, runs the edge model to score each frame, and selects the highest‑scoring image as the cover, using batch asynchronous computation for speed.

Algorithm Research : Various NR‑IQA models were surveyed, including Faster‑VQA, UNIQA, LAR‑IQA, CLIP‑IQA, and Q‑Align. Lightweight MobileNetV3 was chosen as the backbone, and a custom loss combining regression and perceptual bias was designed. Over 100k labeled images were generated via large‑model pre‑annotation and human verification.

Model Migration : The PyTorch MobileNetV3 model is exported to ONNX, then converted to MNN with FP16/Int8 quantization, resulting in a 24 MB model suitable for mobile devices.

Inference Consistency : Consistency across hardware (CPU/GPU), frameworks (PyTorch, ONNX, MNN), and platforms (iOS, Android) is discussed, with emphasis on unified preprocessing and tolerance for minor variations.

Latency Optimization : Parallel frame extraction, GPU inference, and adaptive frame sampling for low‑end phones reduce end‑to‑end latency to sub‑second levels.

Evaluation : Offline and online experiments show that intelligent covers improve visual quality, achieve a 41.7 % GSB score increase, raise smart‑cover CTR by 5.5 %, and boost overall video CTR (PVCTR +13.12 %, UVCTR +18.05 %).

Conclusion : Deploying edge AI for cover recommendation lowers creation cost and enhances user experience, with potential extensions to other scenarios.