Multi-Scale Fire Detection Algorithm in Low-Light Backgrounds

To tackle the challenge of inadequate detection accuracy in intricate settings, particularly due to factors like dim outdoor lighting and obstructed backgrounds, this study introduces a multi-scale fire detection algorithm rooted in the YOLOv10 framework. The proposed solution incorporates a multi-channel convolutional design within the feature extraction network, aimed at bolstering the model's capacity to identify flame targets across various scales. The CARAFE (Content-Aware ReAssembly of FEatures) upsampling operator combines contextual data from broader receptive fields, facilitating the integration of minor flame features. Additionally, an improved EIOU Loss function is employed for fine regression of bounding boxes. The proposed algorithm achieves a detection accuracy of 82.3% on a custom dataset, demonstrating strong anti-interference capabilities while balancing efficiency and real-time performance. This renders it highly effective for fire detection in intricate situations.