Remote Sensing Image Captioning Based on Multi-Level Feature Extraction and Adaptive Attention

Accurate depiction of high spatial resolution remote sensing images necessitates an understanding of the internal attributes of objects and the external relationships between different entities. Existing image captioning algorithms lack global representational capability, rendering them unsuitable for the summarization of intricate scenes. To address this, we propose a remote sensing image captioning approach based on multi-level feature extraction and an adaptive attention mechanism. Specifically, in the encoder phase, localized and global features of the image are extracted separately, and subsequently, the attention mechanism method integrates the global features with locally extracted features from the target detection algorithm. In the decoder section, a third layer LSTM network is introduced as a contemplation layer, concurrently incorporating an adaptive attention mechanism based on image sentinels, enhancing the precision and delicacy of the text description generated by the decoder. Experimental results on public datasets substantiate the superiority of the proposed method over existing approaches.