A dataset named MSIFSD, which contained 9 types of fire scenes and 3 types of smoke, was constructed to explore a lightweight fire smoke detection algorithm with good performance. Additionally, a deep convolutional neural network named DeepSmoke was developed. To address the challenge of detecting small-scale and sparse smoke, an efficient feature aggregation module called PM-C2f was proposed and incorporated with a partial mixed-relevance region self-attention mechanism module called PM-TF. The PM-C2f module was utilized to integrate contextual information from various levels of image features, while the PM-TF module was used to strengthen the sparse features of small-scale and sparse smoke. A semi-supervised training strategy using a pseudo-label classifier was proposed to address the issue of DeepSmoke’s insufficient adaptability across different scenarios. A large amount of unlabeled data were leveraged to assist model training and improve the detection performance across multiple scene types. Experimental results demonstrated that the proposed algorithm achieved the detection accuracies of 88.2%, 90.0%, and 98.3% for small-scale, sparse, and general smoke, respectively. The average comprehensive detection accuracy was 94.2%, outperforming the existing mainstream algorithms. And the floating-point operation was 9.3×10⁹, reflecting the friendliness to edge devices.