Analisis Efisiensi CNN Ringan untuk Retrieval Gambar Tekstur Batu Berbasis Fitur Embedding

Abstrak

Content-Based Image Retrieval (CBIR) is vital for geological applications, yet existing deep learning models often incur high computational costs when processing static 2D stone textures. This study proposes a lightweight Convolutional Neural Network (CNN) architecture optimized for feature embedding extraction to enhance retrieval efficiency without compromising accuracy. We compare the proposed lightweight model against standard CNNs and hybrid architectures using a dataset of 2,118 thin-section rock images across 10 mineral categories. The evaluation focuses on retrieval precision, recall, Mean Average Precision (MAP), and computational complexity including parameters and FLOPs. Results indicate that the lightweight architecture achieves a MAP of 0.74 while reducing model size by 40% compared to baseline models. This demonstrates that specialized, efficient feature extractors are more suitable for scalable geological image retrieval systems than complex hybrid models. The findings provide a practical framework for deploying CBIR in resource-constrained environments such as field surveys and mobile geological applications, offering a balance between high accuracy and low latency.

Identifier

• OpenAlex: https://openalex.org/W7171367487
• DOI: https://doi.org/10.47002/metik.v10i1.1427