Bird sounds have diverse features， and most of the current convolutional neural network models based on a single receptive field are difficult to learn the diversity of bird sound features from audio containing complex background noise. In this article， we proposed a method of classifying bird sounds based on the Inception-CSA deep learning model， which consists of three steps including bird audio sample preprocessing， feature extraction， and classifier classification. First， the samples of bird sounds were preprocessed into Mel spectrum maps with the same size as the feature maps of bird sounds. Then the feature of bird sounds was extracted with the Inception-CSA model including the Inception module extracting the multi-scale local time-frequency domain features in the feature map of bird sounds and the CSA module obtaining the global attention weights of the feature map of bird sounds. The output of both was combined to obtain a stronger feature map. The feature maps were downsampled with the maximum pooling layer. Finally， the results of final classification were obtained with the fully connected layer. The calls of 10 wild bird species in the natural environment of south China were collected and the dataset was constructed to verify the effectiveness of the method. The results showed that the proposed method achieved 93.11% accuracy on the self-built dataset. The classification method based on the Inception-CSA model had higher accuracy with fewer model parameters compared with the classification methods based on other classical models.