Real-Time Pattern Recognition in Augmented Reality Application Through the Integration of TensorFlow Lite and Unity

The integration of machine learning with augmented reality (AR) is vital for traditional batik preservation, especially in real-time pattern recognition. Traditional AR applications often rely on external servers, increasing latency and reducing interaction fluidity. This paper presents real-time batik pattern recognition using quantized TensorFlow Lite (TFLite) in a mobile AR application. Using the Single Shot MultiBox Detector (SSD) architecture with quantization, model size is reduced by 75%, while maintaining 96% accuracy, 0.06 seconds inference time, and 46 frames per second (FPS). Compared to YOLO, which achieves 93% accuracy and 40 FPS, TFLite offers faster inference and reduced computational load, making it ideal for mobile AR. By embedding TFLite into an AR platform, users can interact with batik patterns in real-time without server reliance. This research shows that quantized TFLite enhances mobile AR performance, balancing speed, accuracy, and efficiency.