基于特征选择和GWO-KELM的鸟声识别算法

针对鸟声识别算法中提取特征单一、分类准确率低等问题，提出一种基于混合特征选择和灰狼算法优化核极限学习机的鸟声识别方法。首先从鸟声数据中提取大规模声学特征集ComParE，其次计算每个特征的F_score并进行排序，然后以广义顺序向前浮动搜索(Generalized Sequential Forward Floating Search, GSFFS)为搜索策略，特征子集在核极限学习机(Kernel Limit Learning Machine, KELM)上十折交叉验证的正确率，作为特征选择标准进行特征选择，得到适用于鸟声识别的特征子集，最后通过灰狼算法(Grey Wolf Optimizer, GWO)选择最优KELM参数识别鸟声。在柏林自然科学博物馆鸟声数据库中进行实验，该方法在60类鸟声识别平均正确率和F1-score达到94.45%和92.29%。结果表明，该方法相较于传统自行设计提取的单一特征集具有更高的识别精度，GWO-KELM模型比网格搜索方式更易找到全局最优值。

Bird sound recognition algorithm based on feature selection and GWO-KELM

To address the problems of single feature extracted and low classification accuracy in bird sound recognition algorithms, a bird sound recognition method based on hybrid feature selection and Gray Wolf algorithm-optimized kernel limit learning machine is proposed. Firstly, the large-scale acoustic feature set ComParE is extracted from bird sound data, secondly, the F_score of each feature is calculated and ranked, then the generalized sequential forward floating search (GSFFS) is used as the search strategy, and the correct rate of feature subsets on the kernel limit learning machine (KELM) with ten-fold cross-validation is used as the feature selection criterion to select the features applicable to bird sound recognition subset, and finally the optimal KELM parameters are selected by the Grey Wolf Optimizer (GWO) to recognize bird sounds. In the experiments conducted in the bird sound database of the Berlin Museum of Natural Sciences, the average correct rate and F1-score of the method reaches 94.45% and 92.29% for 60 types of bird sounds. The results show that the method has higher recognition accuracy than the traditional self-designed single feature set, and the GWO-KELM model is easier to find the global optimal value than the grid search method.