Özet:

Abstract The historical temples in India belonged to illustrious kingdoms that ruled for nearly a millennium. Their pillar style, sculpture, inbuilt, architecture, technique, vastness and magnitude have very awesome wonders of their own. Using useful information from the on-site diagnostic of the raw history of the pillar architecture, archaeologists can make decisions about many aspects of pillar handling and management techniques. Archaeologists can better understand old temples by segmenting the pillars, which is useful for future research like identification & classification of different types of pillars based on architecture, to know the original architecture adopted during the construction of temples by various dynasties, and this original architecture information guides the re-construction of temples. Because there are no reliable digital methods for automatic pillar segmentation, archaeologists must deal with a number of challenging issues. Due to irregularities in image acquisition, complex architectural designs, noise, time, and imaging distortions, automated pillar segmentation presents difficulties. In the literature, certain inaccurate statistical segmentation techniques for pillar segmentation have been suggested. For the auto-segmentation of pillars, we suggest a fully convolutional network(FCN) Model in this paper. The suggested technique reduces the unpredictability of picture noise and develops FCN models using images from our own generated dataset. Furthermore, optimal data augmentation and model hyperparametrization are shown to prevent overfitting for pillar area segmentation. With a recall/precision rate of 0.9698/0.9200, the proposed approach is examined on the test dataset. When compared to published algorithms in the literature segmentation challenge, the new method performs better, with a Dice correlation coefficient of 0.9284, than those algorithms.

Anahtar Kelimeler: