Spatial Attention Deep Convolution Neural Network for Call Recognition of Marine Mammal

Abstract

Aiming at the task of call recognition of marine mammal, the time-frequency effective feature spatial focus method is proposed. The method tends to use Spatial Attention (SA) to help Deep Convolution Neural Network (DCNN) reach better recognition performance. The Time-Frequency Image Recognition_DCNN (TFIR_DCNN) of the method is designed first. The proposed network realizes the recognition task with time-frequency images of the calls as input. Features of time-frequency image contain the noise features and the call features. Then, to help TFIR_DCNN focus on the spatial position of the call features in time and frequency domain, the SA is added to TFIR_DCNN. SA can generate and multiply weight tensor to the feature maps so the features of the call are emphasized on spatial to help TFIR_DCNN recognizing. The paper designs a call recognition of marine mammal experiment. In the experiment, model from the proposed method achieves 66.14% in recall, 73.55% in precision, 66.14% in accuracy, 69.95% in F1-score and gets 5.29% in recall, 10.01% in precision, 5.29% in accuracy, 7.48% in F1-score higher than the control model.