Mobile AI Digit Recognition for Game Live Streaming: Model Training, Optimization, and Cross‑Platform Deployment

This article describes the development of a mobile AI digit‑recognition system designed to identify kill counts and remaining player numbers in the popular battle‑royale game “PUBG Mobile” during live streaming.

Background: Traditional deep‑neural‑network inference runs on servers, but modern mobile SoCs (Apple A11/A12, Qualcomm Snapdragon 845/710, Huawei Kirin 970/980) include dedicated AI accelerators, enabling on‑device inference with frameworks such as TensorFlow Lite or CoreML.

Data Collection & Pre‑processing: The training set combines the MNIST handwritten digit dataset (250 writers, balanced high‑school students and census workers) with a custom‑collected printed‑digit dataset obtained via web crawling and manual capture. Images are binarized, noise‑filtered, and converted to the MNIST format.

Model Choice & Adjustments: A simple CNN (two convolution‑pooling layers and two fully‑connected layers) is selected for its ease of understanding and >99% accuracy on MNIST. To suit mobile constraints, the large fully‑connected layer is removed, an extra convolution‑pooling layer (average pooling) is added, reducing model size from ~13 MB to <500 KB and RAM usage from ~10 MB to ~5 MB.

Performance Evaluation: On Ubuntu 18.04 with TensorFlow‑GPU 1.2 and GTX 1060, the optimized model retains >99% accuracy, with comparable convergence speed. Mobile tests on iPhone 8 (iOS) and Xiaomi Mix 2S (Android) show inference latency <10 ms and CPU usage <5%.

iOS Deployment: The TensorFlow .pb model is converted to Apple’s CoreML format using the tf‑coreml Python library. The CoreML model is integrated into a Swift‑based live‑streaming app, leveraging dynamic compilation and built‑in classification APIs.

Android Deployment: The same .pb model is converted to TensorFlow Lite (.tflite) format. The Android app extracts digit regions from the game screen, preprocesses them similarly, and runs inference via the TFLite interpreter.

Cross‑Platform Integration: Both platforms send prediction results to a backend server via HTTP, where the data is aggregated and displayed to viewers.

Conclusion: Mobile AI enables real‑time digit recognition without additional server infrastructure, achieving high accuracy and low resource consumption. Future work includes expanding AI‑driven features and further model optimization for newer devices.