Optimization of CNN Activation Functions using Xception for South Sulawesi Batik Classification

Abstract

Batik motifs from South Sulawesi such as the Pinisi boat, Lontara script, Tongkonan house and Toraja combinations embody rich cultural narratives but are difficult to identify automatically. Automatic classification supports cultural preservation and education and empowers tourism and digital heritage applications. This study improves the performance of convolutional neural networks for South Sulawesi batik classification by optimizing activation functions within the Xception architecture which exploits depthwise separable convolutions for efficient and detailed feature extraction. A balanced dataset of 1400 labeled images in four classes was divided into eighty percent for training, ten percent for validation and ten percent for testing. Images were resized to 224 by 224 pixels, converted to grayscale and augmented through zoom, flip and rotation. With identical hyperparameters including a learning rate of 0.001, a batch size of 64 and training for 100 epochs using the Adam optimizer, ReLU, ELU, Leaky ReLU and Swish activation functions were compared. Evaluation metrics comprised accuracy, precision, recall, F1 score and cross entropy loss. ELU achieved the highest test accuracy of 98.57 percent, precision of 0.9864, recall of 0.9857 and F1 score of 0.9857, outperforming ReLU and Leaky ReLU with 97.86 percent accuracy and Swish with 97.14 percent accuracy. The results demonstrate that selecting an optimal activation function substantially enhances convolutional neural network classification of complex batik patterns. The findings offer practical guidance for development of resource aware batik identification systems in support of cultural digitization and education initiatives.