Early smoke detection is an effective means to eliminate fire hazards in a timely manner, but the small size and complex diffusion form of smoke in the early stage of a fire make its detection extremely difficult. To address the above problems, this paper proposes a multi-path enhanced feature-based YOLO （MEF-YOLO）early smoke detection algorithm, which adopts QA-ELAN to improve the backbone network and optimise the model complexity and accuracy, and develops FGCA to autonomously enhance the feature differences between the sampling regions to effectively capture the spatial information of the smoke. And the feature fusion path is optimised by the MEFAN, which realises the direct interaction between cross-level features and effectively mitigates the loss of detail information; and a Wise-IOU loss function is introduced, which comprehensively takes into account the position and scaling information through the weight adjustment mechanism to improve the robustness of the model in the complex scene. The experimental results show that the algorithm proposed in this paper has an accuracy of up to 92.5% for early smoke detection in experimental scenarios with different lighting and small-scale smoke and smoke diffusion, and has a lightweight advantage, with the number of parameters and GFLOPs reduced by 27.5% and 30.6%, respectively.