FPGA-Based Batik Classification Using Quantization Aware Training of MobileNet and Data-flow Implementation

Batik, a traditional Indonesian textile recognized by UNESCO, poses unique challenges in pattern recognition due to its intricate and varied designs. While convolutional neural networks (CNNs) have shown promise in Batik classification, they require substantial computational resources and suffer from limitations in local feature detection. This paper proposes a novel approach leveraging Quantization Aware Training MobileNet (QAT-MobileNet) to classify Batik patterns efficiently on FPGA hardware. Our approach significantly reduces computational complexity through quantization without sacrificing accuracy, achieving superior performance over traditional CNN models. We demonstrate the effectiveness of QAT-MobileNet using the Batik Nitik 960 dataset, achieving a $99.1 \%$ accuracy and competitive FPGA hardware performance with a reduced power footprint. This work not only advances Batik classification but also illustrates the potential of quantized deep learning models for hardware-efficient implementations.