Abstract: In response to the challenge of balancing accuracy and real-time performance when deploying flame detection models on edge computing platforms for visual security systems,a lightweight flame detection algorithm featuring progressive adaptive feature fusion is proposed.Firstly,a lightweight sparse convolution operator is designed to reduce the model’s computational complexity and memory access cost.Subsequently,to address the shortcomings of inter-channel information exchange in grouped convolutions,a lightweight feature extraction component is constructed based on the residual concept,enhancing long-distance contextual features.To tackle the issues of feature loss and background interference in deep backbone networks,an innovative lightweight feature enhancement mechanism based on high-frequency augmentation is proposed,optimizing the parameters in both spatial and channel domains to mitigate background disturbances.On this basis,a feature enhancement-progressive adaptive feature fusion framework is established to facilitate the thorough integration of feature maps at different scales,thereby improving the utilization of feature maps and enhancing the recognition effectiveness of multi-scale targets.Experimental results demonstrate that this method achieves a real-time inference speed of up to 27.1 FPS,reduces the parameter count to 2.1 M,which is a 69.5% reduction compared to the baseline model,and attains a detection accuracy of 83.4% mAP@0.5,significantly outperforming existing mainstream methods.

Key words: Deep learning, Computer vision, Object detection, Lightweight neural network, Feature extraction network, Feature fusion network, Feature enhancement