Abstract: Early fire detection in high-rack storages is challenging due to dense cargo storage, complex structural configurations, and lighting interference. To address this issue, a lightweight deep learning algorithm that fuses visible and infrared images is proposed. The method employs a dual-stream feature extraction network and incorporates a cross-modal feature complementation mechanism to effectively enhance the recognition capabi-lity for smoke and flame characteristics. Additionally, a DFASC-IPV auxiliary module based on video grayscale variation analysis is designed to effectively suppress light source interference and enhance responsiveness to thin smoke. Experimental results demonstrate that compared with single-modal methods, the proposed approach achieves an accuracy improvement of 11.46% and a recall improvement of 13.14% in complex storage scenarios, exhibiting excellent robustness and practicality. This research provides a reliable technical solution for early fire detection in high-rack storages.

Key words: warehouse fire detection, bimodal fusion, attention mechanism, DFASC-IPV