<![CDATA[The development of information technology can be utilized to support the preservation of traditional cultural heritage, including Balinese songket, which possesses diverse motifs and high cultural value. However, the influence of modernization has reduced younger generations’ understanding of Balinese songket motifs. Based on a survey involving 21 respondents aged 20–29 years, 47.6% showed low understanding of Balinese songket motifs, 23.8% had moderate understanding, and 23.8% did not understand the motifs at all. This condition indicates the existence of a cultural knowledge gap that requires preservation efforts through digital technology. Therefore, this study aims to develop a Cross-Modal Retrieval system for Balinese songket motif recognition using a Deep Learning approach.The proposed system utilizes a combination of Convolutional Neural Network (CNN) and Bidirectional Encoder Representations from Transformers (BERT) to model the semantic relationship between visual and textual data. The dataset consisted of 108 image-caption pairs divided into 35 Balinese songket motifs. In this study, ResNet was employed as a visual feature extractor to capture motif patterns, textures, and color characteristics from images, while BERT was used to generate contextual textual embeddings from caption descriptions. Both visual and textual embeddings were projected into a shared embedding space to enable text-to-image and image-to-text retrieval through similarity matching.Prior to training, image preprocessing was performed through cropping, resizing, and image augmentation techniques such as flipping and rotation to improve data variability. Text preprocessing included lowercasing and tokenization to standardize textual input. The model was trained for 300 epochs using a batch size of 16, learning rate of 3e-5, and AdamW optimizer. Experimental results obtained a train loss of 0.4165, validation loss of 2.0716, and Recall@K of 91.67%. These results indicate that the proposed model successfully generated discriminative embedding representations and achieved high retrieval accuracy in matching relevant image-text pairs.Although the model achieved strong retrieval performance, the gap between train loss and validation loss indicates limitations in generalizing to unseen data. This issue is influenced by the high visual similarity among several Balinese songket motifs, particularly in geometric patterns, woven textures, and color compositions, which reduce inter-class embedding separation. Nevertheless, the experimental results demonstrate that the Cross-Modal Retrieval approach effectively integrates visual and textual information for Balinese songket classification and retrieval tasks.In conclusion, the proposed system shows strong potential as an interactive digital medium for preserving and introducing Balinese songket motifs, especially for younger generations. Future work can focus on expanding the dataset, improving semantic caption quality, applying more diverse augmentation strategies, and optimizing fine-tuning, feature fusion, and attention mechanisms to enhance embedding quality and reduce the semantic gap between image and text modalities.]]>
