Automated Multi-Class Classification of Batik Klowong Defects Using CNNs: A Comparative Study of Transfer and Full Learning Approaches

Abstract

Batik, an internationally recognized cultural heritage of Indonesia, requires rigorous quality control to preserve both its artistic integrity and commercial value. Conventional manual inspection of Batik klowong defects is labor-intensive, time-consuming, and prone to human error, so an automated, reliable alternative is needed. This study proposes a Convolutional Neural Network (CNN)-based framework that formulates Batik klowong defect detection as a domain-specific multi-class classification problem. The dataset was collected during the klowong process and comprises 6,072 images categorized into eight defect classes and one non-defective class. The framework integrates transfer learning with a custom-designed CNN architecture and evaluates eight models, EfficientNet-B0, MobileNetV2, DenseNet121, ResNet50, VGG19-BN, SqueezeNet, ShuffleNetV2, and a fully trained CNN, under a unified protocol. We assessed model performance using accuracy, precision, recall, and F1-score. VGG19-BN achieved the highest accuracy (99.11%), followed by DenseNet121 (98.44%) and MobileNetV2 (98.33%). EfficientNet-B0 reached 97.66% while offering improved computational efficiency. Lightweight models such as SqueezeNet and ShuffleNetV2 showed lower reliability in capturing complex defect patterns. The study further incorporates uncertainty-aware statistical evaluation using Wilson confidence intervals and Cohen’s h effect size to ensure more reliable model comparison. Overall, VGG19-BN and DenseNet121 provide robust defect classification, whereas MobileNetV2 and EfficientNet-B0 demonstrate strong real-time feasibility for resource-constrained industrial deployment. Although we did not explicitly measure inference latency, future work should evaluate real-time performance and optimize lightweight architectures for industrial applications.