Early fire detection is crucial for reducing losses due to fire. Therefore, the development of an efficient fire detection algorithm and alarm system is essential. The image fire detection algorithm is based on the mathematical analysis of images. Current algorithm evaluation methods are not effective and cannot sufficiently distinguish the performance of different detection algorithms. Therefore, such methods are not conducive for evaluating and improving algorithms. Based on a large-scale fire image dataset, the ground-truth complexity of images was quantified according to the time required for humans to detect the presence or absence of fire in the images. Four image complexity metrics based on the characteristics of fire detection are proposed. A comparison of the ground-truth and predicted scores of image complexity revealed that the comprehensive image complexity metric based on the Inception Resnet -v2 predictor was the most effective measurement. Its predicted scores ranked approximately 85% image pairs in the same order as that of the ground-truth complexity scores. Finally, a novel method for evaluating the performance of an image fire detection algorithm based on image complexity is proposed. Evaluation of the performance of five algorithms revealed that the performance of algorithm differs considerably and the proposed method can accurately determine the detection level of the detection algorithm in different image complexity conditions. The results of the study provide a valuable reference for developing and optimizing detection algorithms.