Fire is one of the most common social disasters in daily life, which will pose an enormous threat to human property and life safety. How to accurately and quickly identify small areas of smoke and fire and issue early warnings in real time is important for normal social production significance. The traditional smoke and fire detection algorithm identifies the location of smoke and fire based on various low-dimensional visual features of the images, such as color and texture, so it is of poor real-time performance and low accuracy. In recent years, deep learning has made remarkable achievements in the field of target detection, and various smoke and fire detection methods based on deep neural networks have sprung up one after another. In the case of small areas of smoke and fire, timely identification and early warning should be made to avoid greater economic losses caused by the expansion of the fire. In this regard,  based on the YOLOX model, the activation function and loss function were improved, and a superior small target detection algorithm was realized by combining the data augmentation algorithm and cross-validation training method, and the mAP value of 78.36% was obtained on the smoke and fire detection data set. Compared with the original model, it was enhanced by 4.2%, yielding a better effect of small target detection effect.