Abstract: In view of the problem that the sample size of highway tunnel fire is small and the deep learning effect is not ideal, a small sample learning technology is studied to improve the full utilization of tunnel fire samples, and on this basis, a fire recognition technology of tunnel video based on self attention is proposed by using mature machine learning methods. The technology uses self-attention mechanism combined with SVM classifier to build a flame recognition model. According to the importance of each feature to flame recognition, the model assigns different attention weights to form the attention matrix, and weights the weight matrix with the value of the feature vector.Hinge Loss of SVM is used for linear support vector machine classification, so as to identify and warn highway tunnel fire.In the process of fire recognition training, through testing the flame suspected area, and use the data to enhance technology to achieve the purpose of the sample amplification, then the multi-modal fusion way of feature extraction was used to construct characteristic vector and input to the self-attention flame recognition model. Through the gradient descent optimizer, small batch model training is conducted, to reduce the number of iterations, Finally, the optimal feature weight parameters and the best recognition model are obtained. The experimental results show that the method converges faster during model training, and in flame identification, compared with the traditional SVM algorithm without small sample learning, the proposed method improves the accuracy by 5% in flame recognition, and can effectively improve the accuracy of fire recognition in a small sample environment.

Key words: tunnel video, video flame identification, small sample learning, self-attention mechanism, the SVM algorithm