earticle

영어

In this paper, we propose a novel on-device AI design strategy for dynamically adjusting subtitle positioning in IPTV content that must be stored within a secure environment to prevent unauthorized distribution. Unlike previous studies that rely on external servers or unsecured environments for video analysis, our approach embeds the AI model directly into the secure zone of the chipset, ensuring privacy and real-time performance. We uniquely utilize on-device hardware resources to enable AI-based frame-by-frame analysis without external transmission. To achieve real-time efficiency, we implement a lightweight and device-optimized AI model by defining a Region of Interest (ROI) for input videos, applying model pruning, and utilizing 8-bit quantization techniques. Additionally, we enhance text recognition performance through data augmentation during training, addressing common challenges such as subtitle overlapping with on-screen graphics or embedded text. We demonstrate that our on-device strategy outperforms conventional models, improving recognition accuracy to 99.7% and processing speed to 60 fps. Through this work, we contribute a practical solution that ensures enhanced subtitle visibility for IPTV viewers using set-top boxes while maintaining content security in a non-trainable trusted execution environment.