Fire disasters pose significant risks to human life, economic development, and social stability. The early stages of a fire, often characterized by small flames, diffuse smoke, and obstructed objects, can lead to challenges such as missed detections and poor real-time performance. To address these issues, we propose a DSS-YOLO model based on an improved YOLOv8n architecture, designed to enhance the recognition accuracy of obscured objects and small targets while reducing computational overhead. Specifically, we replace all C2f modules in the Backbone with DynamicConv modules to reduce computation without sacrificing feature extraction capabilities. We also introduce the SEAM attention mechanism to improve detection of obscured and small targets, and the SPPELAN module at the end of the Backbone to enhance detection across different scales. The model is evaluated using the public dataset mytest-hrswj, which contains diverse fire scenarios, including indoors, forests, and buildings. Compared with the original YOLOv8n, the DSS-YOLOv8 model proposed in this paper improves mAP by 0.6% and Recall by 1.6%, while reducing the model size and FLOPs by 3.4% and 12.3% respectively. The results of this study provide effective technical support for intelligent fire monitoring systems, significantly reducing the computational cost of the model. It enhances real-time fire detection capabilities in complex fire scenarios, facilitating the early detection of fire hazards and helping to minimize the damage caused by fires.