Interpreting Geometric Constructions in Artworks through Capsule Network Modeling

Abstract

Abstract Interpreting the geometric structure of artworks enhances our intuitive grasp of their deeper meanings. This study employs a Capsule network model, incorporating a dynamic routing algorithm to correlate high and low-level geometric structural features of artworks. Additionally, an attention mechanism is introduced, forming a spatial attention capsule to capture the spatial context of the artwork’s geometric structure. To obtain images, a fixed-focus camera is utilized, followed by median filtering for image preprocessing and threshold segmentation using the maximum inter-class variance method to optimize recognition accuracy. The efficacy of the geometric structure recognition model, grounded in the Capsule network, is confirmed using a dataset of collected artwork images. The model achieves stability after 380 epochs, exhibiting an impressive accuracy of approximately 99.7% and a minimal loss of 0.025. Removing the attention mechanism results in a 4.06 percentage point decrease in model accuracy, whereas incorporating a dynamic routing algorithm boosts efficiency by 7.36%. Thus, the Capsule model proves highly effective in precisely recognizing and interpreting the geometric structures of artworks.