Abstract: An improved recognition algorithm was constructed by deeply integrating multi-task convolutional neural network (MTCNN) detection module, octave convolution (OctConv) feature optimization module and FaceNet recognition module. OctConv was introduced to replace the traditional convolution of MTCNN, a loss function for keypoint regression was added to output network (O-Net), a frequency-domain dynamic weight allocation mechanism was introduced to enhance OctConv module, and the triplet loss function was added during FaceNet training phase. Meanwhile, a novel facial image dataset covering various typical thermal power plant scenarios was constructed. Experimental results show that the accuracy of the proposed optimized algorithm reaches 95.2%. It performs excellently in the high-temperature area of the boiler, with the recognition accuracy increased from 87.6% of traditional MTCNN algorithm to 97.2%, effectively solving the problem of feature degradation caused by the reflection of high-temperature equipment. The detection speed of this algorithm reaches 24.7 frames per second, which is superior to that of MTCNN (22.3 frames per second) and convolutional neural network (CNN)-long short-term memory (LSTM) network (18.5 frames per second). The optimized algorithm improves accuracy and real-time performance, reduces computing resource consumption, can effectively meet the safety management requirements in the complex environment of thermal power plants, and enhances the intelligent level of safety management in thermal power plants.