KD-CLDNN: Lightweight automatic recognition model based on bird vocalization

Abstract

Passive acoustic monitoring (PAM) equipment embedded automatic bird recognition is conducive to the real-time monitoring of birds. In this paper, the lightweight bird recognition model KD-CLDNN is proposed to adapt to the monitoring equipment with limited computing power. The KD-CLDNN model is obtained by knowledge distillation. The teacher model CS-CLDNN (CBAM-Switch-CLDNN) is built by introducing Convolutional Block Attention Module (CBAM) and Swish activation function to CLDNN model. Furthermore, the student model Net-S is constructed by simplifying the CS-CLDNN model. The recordings of 20 bird species are used as the dataset in this study, and the performances of six complex models are evaluated with this dataset. The results show that the CS-CLDNN model outperforms other complex models. Compared with the CS-CLDNN model, the performance of the KD-CLDNN model in accuracy, recall, precision, and F1 score decrease slightly by 0.028, 0.025, 0.01, and 0.023, respectively. However, the training GPU time of KD-CLDNN is reduced about seven times, and the total number of parameters is reduced about three times. On the edge computing platform NVIDIA Jetson TX2, the inference speed of KD-CLDNN is 3.2 times faster than that of CS-CLDNN. Therefore, the lightweight KD-CLDNN model can significantly reduce calculation requirement on the premise of ensuring recognition accuracy, which can be helpful for the development of intelligent bird monitoring equipment based on deep learning.