YOLOv8n-AM: Enhanced Real-Time Smoke Detection via Attention-Based Feature Interaction and Multi-Scale Downsampling

Accurate smoke detection in complex industrial environments, such as chemical plants, remains a significant challenge due to the inherently low contrast, transparency, and weak texture features of smoke, which often exhibits blurred boundaries and diverse spatial scales.To address these limitations, YOLOv8n-AM, an enhanced lightweight detection framework belonging to the YOLO (You Only Look Once) series, was developed by integrating advanced architectural components into the baseline YOLOv8n model.Specifically, the conventional Spatial Pyramid Pooling-Fast (SPPF) module was replaced with an Attention-based Intra-scale Feature Interaction (AIFI) Convolution Synergistic Feature Processing Module (SFPM), i.e., AIFC-SFPM, enabling more effective semantic feature representation and an improvement in detection accuracy.In parallel, the original convolutional module was optimized using a Multi-Scale Downsampling (MSDown) module, which reduces model redundancy and computational overhead, increasing the detection speed.Experimental evaluations demonstrate that the YOLOv8n-AM model achieves a 1.7% improvement in mean Average Precision (mAP), accompanied by a 9.1% reduction in Giga Floating-point Operations Per Second (GFLOPs) and a 15.4% decrease in parameter count when compared to the original YOLOv8n framework.These improvements collectively underscore the model's suitability for real-time deployment in resource-constrained industrial settings where rapid and reliable smoke detection is critical.The proposed architecture thus provides a computationally efficient and high-precision solution for safety-critical visual monitoring applications.