Classification of birdsong spectrograms based on DR-ACGAN and dynamic convolution

Birdsongs are highly valuable for bird studies as they provide insights into various aspects such as species distribution, population structures, and habitat. Recognizing birdsongs plays a crucial role in bird conservation efforts. However, manually collecting a large number of birdsongs from the natural environment is expensive and time-consuming. Moreover, using limited birdsong data often results in low classification accuracy of the models. To better identification of birdsongs, we utilize wavelet transform(WT) to convert them into spectrograms, which contain abundant energy and frequency information. Effectively extracting these features is vital to improve the classification accuracy of the model. To address this problem, we proposed an improved ACGAN model based on residual structure and attention mechanism named DR-ACGAN, which achieved stable training of the model and high-quality generated birdsong spectrograms. The dynamic convolution kernel is then fused with MobileNetV2, ResNet18, and VGG16 models and trained on different datasets, which used different ways of mixing the generated and original spectrograms. The experimental results show that the classification accuracy after data augmentation improves by 6.66%, 4.35%, and 2.29% compared to the original dataset in the three base classifiers. After adding dynamic convolutional kernel structure, the accuracy is further improved by 1.68%, 0.67%, and 0.38% on average which the VGG16 model achieves the highest accuracy of 97.60%.