MSFAEGD: Efficient Forest Smoke Detection Network Based on Grounding Dino

Forest smoke and fire pose a significant threat to global ecological security, with prevention and control efforts hampered by the complexity and multifaceted challenges associated with technological methods and resource allocation. Existing forest smoke and fire detection techniques exhibit notable limitations, including narrow monitoring ranges, high false alarm rates, and insufficient accuracy in smoke and fire recognition. In this paper, we propose a two-stage forest smoke and fire detection model based on Grounding DINO, named MSFAEGD. The model innovatively designs an attention-guided multiscale feature aggregation and extraction module, AMSFAE, for the multiscale characteristics and complex spatial distribution of forest smoke and fire. The AMSFAE consists of sequentially stacked scale and spatial position attention components, which not only efficiently capture rich multiscale features, but also accurately localize salient regions in the image space. Experimental results demonstrate that MSFAEGD, finetuned on a forest smoke and fire dataset, not only outperforms the original Grounding DINO but also significantly surpasses most existing state-of-the-art methods with comparable parameters, achieving an impressive AP of 63.2.