Abstract: The "image complexity" is proposed to measure the difficulty for the algorithm to detect fire in an image. The "human response time" is obtained through the program test method, which could define the true value of the complexity of the image in the data set. Then, convolutional neural network is used to develop an evaluator that can automatically predict the comprehensive complexity of the whole image. Through the consistency analysis, the comprehensive complexity measure of the whole image generated by the Inception ResNet-v2 evaluator is the most effective measure. Furthermore, a performance evaluation method based on image complexity is proposed. This method can evaluate the detection level that the algorithm can achieve under different "image complexity" situations more accurately, providing more valuable reference for the development and optimization of fire detection algorithms.

Key words: image complexity, fire detection, convolutional neural network