To address the low detection accuracy and slow speed of traditional flame detection model, a video flame region detection method based on optimal convolutional neural network and hyperpixel segmentation algorithm is proposed. First, the flame image dataset is used to train and verify the model, and the structure of the Inception module is improved by stacking and replacing the convolution kernel. Second, the small convolution kernel replacement is adopted to improve the front-end structure of the network, and the Focal-Loss function is used as the loss function to improve the generalization ability of the model. Next, the parameter complexity optimization experiment of the InceptionV1 model is designed to generate an optimized flame detection network structure. Finally, the flame super-pixel semantic information extracted by the superpixel segmentation algorithm is input into the optimized InceptionV1 model, and the location detection of the video flame area is further performed. Experimental results show that the proposed method can enhance the nonlinear feature extraction of video flames; the accuracy of flame detection is higher than 96%, and the detection speed is 2.66 times that of the original model.