RESUMO
Deep convolutional neural networks (DCNNs) have achieved breakthrough performance on bird species identification using a spectrogram of bird vocalization. Aiming at the imbalance of the bird vocalization dataset, a single feature identification model (SFIM) with residual blocks and modified, weighted, cross-entropy function was proposed. To further improve the identification accuracy, two multi-channel fusion methods were built with three SFIMs. One of these fused the outputs of the feature extraction parts of three SFIMs (feature fusion mode), the other fused the outputs of the classifiers of three SFIMs (result fusion mode). The SFIMs were trained with three different kinds of spectrograms, which were calculated through short-time Fourier transform, mel-frequency cepstrum transform and chirplet transform, respectively. To overcome the shortage of the huge number of trainable model parameters, transfer learning was used in the multi-channel models. Using our own vocalization dataset as a sample set, it is found that the result fusion mode model outperforms the other proposed models, the best mean average precision (MAP) reaches 0.914. Choosing three durations of spectrograms, 100 ms, 300 ms and 500 ms for comparison, the results reveal that the 300 ms duration is the best for our own dataset. The duration is suggested to be determined based on the duration distribution of bird syllables. As for the performance with the training dataset of BirdCLEF2019, the highest classification mean average precision (cmAP) reached 0.135, which means the proposed model has certain generalization ability.
RESUMO
Recognizing wildlife based on camera trap images is challenging due to the complexity of the wild environment. Deep learning is an optional approach to solve this problem. However, the backgrounds of images captured from the same infrared camera trap are rather similar, and shortcut learning of recognition models occurs, resulting in reduced generality and poor recognition model performance. Therefore, this paper proposes a data augmentation strategy that integrates image synthesis (IS) and regional background suppression (RBS) to enrich the background scene and suppress the existing background information. This strategy alleviates the model's focus on the background, guiding it to focus on the wildlife in order to improve the model's generality, resulting in better recognition performance. Furthermore, to offer a lightweight recognition model for deep learning-based real-time wildlife monitoring on edge devices, we develop a model compression strategy that combines adaptive pruning and knowledge distillation. Specifically, a student model is built using a genetic algorithm-based pruning technique and adaptive batch normalization (GA-ABN). A mean square error (MSE) loss-based knowledge distillation method is then used to fine-tune the student model so as to generate a lightweight recognition model. The produced lightweight model can reduce the computational effort of wildlife recognition with only a 4.73% loss in accuracy. Extensive experiments have demonstrated the advantages of our method, which is beneficial for real-time wildlife monitoring with edge intelligence.