A Comparative Study of Deep Learning Architectures for Content-Based Image Retrieval on a Stone Texture Dataset

Abstract

Purpose: This study addresses a persistent gap in geological content-based image retrieval (CBIR): the reliance of existing sequence-based and hybrid deep learning models on static, two-dimensional rock texture representations that they were not originally designed to handle.

Methods: Four architectures were evaluated on a unified 9,853-image, ten-category Stone-2D dataset, partitioned into training and validation subsets at an 8:2 ratio (7,882/1,971 images), with FAISS employed for similarity search over 128-dimensional embeddings.

Findings: On an identical held-out test set, all four architectures achieved comparable performance (82.4%–85.2% accuracy; macro F1 0.826–0.860), with a standard 2D CNN achieving the highest accuracy (85.2%) and the custom pseudo-3D MineralNet architecture achieving the highest macro F1 (0.860). No architecture dominated across both metrics.

Novelty: These results indicate that, once preprocessing is aligned with the native dimensionality of the image data, architectural complexity provides limited additional benefit for static geological texture classification — challenging the common assumption that specialized or sequential architectures are necessary for this task. A working CBIR retrieval pipeline was validated with representative query examples, demonstrating semantically consistent nearest-neighbour retrieval across distinct rock textures.