Abstract:

In order to recognize the fire image quickly and effectively, a flame recognition method of Twin Support Vector Machine（TWSVM） based on Improved Artificial Fish Swarm Algorithm（IAFSA） is proposed. Firstly, this method segments the flame image according to the distribution characteristics of flame pixels in RGB-YCbCr mixed color space model, and extracts the relevant features of the flame image on this basis. Secondly, Artificial Fish Swarm Algorithm（AFSA） is used to search the optimal penalty parameter and kernel parameter of TWSVM. In AFSA algorithm, a clustering-based fish initialization method is used to obtain uniform initial fish swarm. At the same time, adaptive parameters are used to adjust the visual range and moving step length of artificial fish swarm. In addition, based on the original three behaviors, two new behaviors are proposed：jumping behavior and eliminating regeneration behavior, which improves the efficiency and accuracy of fish swarm algorithm. Then the extracted flame features are input into TWSVM model as training samples for training. Finally, the samples to be tested are input into TWSVM model for classification and recognition. Experimental results show that compared with VGGNet model of deep convolution neural network, Fast R-CNN algorithm, YOLO algorithm, traditional Support Vector Machine（SVM）, Grid-TWSVM, GA-TWSVM, PSO-TWSVM, FOA-TWSVM, GSO-TWSVM, AFSA-TWSVM, the proposed method of twin support vector machine based on improved artificial fish swarm algorithm effectively improves the accuracy and real-time performance of flame recognition, and solves the problems of TWSVM such as difficult parameter selection in flame recognition and long optimization time of common parameter optimization algorithms.

Key words: twin support vector machine, improved artificial fish swarm algorithm, flame recognition, parameter optimization, RGB-YCbCr mixed color space model

摘要：

为了快速有效地识别火灾火焰图像，提出了一种基于改进人工鱼群算法（IAFSA）的孪生支持向量机（TWSVM）的火焰识别方法。该方法根据RGB-YCbCr混合颜色空间模型中火焰像素的分布特点对火焰图像进行分割，并在此基础上提取火焰图像的相关特征；采用人工鱼群算法（AFSA）搜索TWSVM最优惩罚参数与核参数，并在AFSA算法中利用基于聚类的鱼群初始化方法来获得均匀的初始鱼群，同时采取自适应参数来调整人工鱼群的视觉范围和移动步长，另外在原有的三种行为的基础上提出了两种新的行为：跳跃行为和淘汰重生行为，提高了鱼群算法的寻优效率和求解精度；将提取的火焰各个特征量作为训练样本输入TWSVM模型进行训练；将待测试样本输入TWSVM模型进行分类识别。实验结果表明：相对于深度卷积神经网络VGGNet模型、Fast R-CNN算法、YOLO算法、传统支持向量机（SVM）、Grid-TWSVM、GA-TWSVM、PSO-TWSVM、FOA-TWSVM、GSO-TWSVM、AFSA-TWSVM，所提出的基于改进人工鱼群算法的孪生支持向量机的方法有效地提高了火焰识别准确率和实时性，解决了TWSVM在火焰识别时参数选择困难、常用参数寻优算法寻优时间长等问题。

关键词: 孪生支持向量机, 改进人工鱼群算法, 火焰识别, 参数优化, RGB-YCbCr混合颜色空间模型