InSwAV: involution enhanced feature clustering and swapped assignments for porcelain relic microscopic image classification

Abstract The precise classification of porcelain relic fragments is a crucial step in the process of artifact restoration and reconstruction. The current classification methods based on apparent features such as glaze color and patterns can achieve preliminary classification, but they have obvious limitations when dealing with porcelain fragments with highly similar visual features. Especially when the sample size of microscopic images is limited, the classification performance of traditional methods often drops significantly. To address the challenges of high computational complexity and low efficiency in traditional contrastive learning for image classification, we propose InSwAV, which aligns deep features with cluster prototypes by swapped assignment through involution-enhanced feature extraction and cross-view consistency enforcement. First, we propose ResInv combining involution blocks with residual blocks of the feature extraction network. Next, features extracted by ResInv are clustered, and consistency in cluster assignments is enforced across multi-view augmented samples via cross-prediction mechanisms. This approach iteratively optimizes feature representation by minimizing cross-entropy loss, significantly reducing training time while maintaining accuracy. Furthermore, we constructed the Porcelain Relic Microscopic Images(PRMI) dataset of five classification, with data augmentation applied to enhance model robustness. Experimental results show that InSwAV achieves a classification accuracy of 96.2% on the porcelain relic microscopic image dataset, outperforming existing methods.