<![CDATA[Convolutional Neural Network (CNN) models achieve high accuracy in visual recognition, but their size and computational cost limit deployment on mobile devices, so batik motif recognition remains difficult to bring to artisans and small creative enterprises. This study evaluates how post-training quantization with TensorFlow Lite affects the classification accuracy, model size, and inference latency of a MobileNetV2 transfer-learning model for batik motif recognition, in order to identify the deployment configuration that offers the best balance among the three. The experimental workflow adapted a pre-trained MobileNetV2 network to the Batik Wastra Nusantara dataset, which covers batik designs originating from different parts of Indonesia with substantial variation in color, ornamentation, and pattern complexity, converted the trained model into Dynamic Range, Float16, and Int8 TensorFlow Lite variants, and compared them under identical conditions. Evaluation on 644 valid images from 14 motif classes produced a test accuracy of 43.08%, corresponding to 28 of the 65 test images, a macro F1-score of 41.65%, and a weighted F1-score of 40.39%. Across the four model formats the accuracy spans only three test images, from 41.54% to 46.15%, a range that lies well inside the confidence interval implied by a 65-image test subset, so the variants cannot be separated on accuracy at this sample size. The deployment metrics differ substantially: relative to the unquantized reference, the Float16 variant reduced model size from 19.04 MB to 4.29 MB (77.47%) and inference latency from 205.52 ms to 28.60 ms (86.08%), while the Int8 variant was the fastest at 14.87 ms and the Dynamic Range variant the smallest at 2.41 MB. These findings indicate that the benefit of TensorFlow Lite optimization lies in storage and latency rather than in accuracy, and that the accuracy attainable on this dataset is constrained by its limited size and the visual similarity among motifs. The study contributes the deployment measurements, namely model size and inference latency for each quantization variant, that comparable batik classification studies do not report, and identifies dataset expansion as the prerequisite for reaching production-level accuracy in AI-based digital innovation for Indonesia’s creative industries.]]>
Copyrights © 2026
