An Attention Pooling based Channel Accumulation for Target Classification of Harmonic Radar

This paper proposes an attention pooling method based on channel accumulation of spectrograms for harmonic radar target classification. The traditional method of audio recognition is feeding the two-dimensional (2D) spectrograms directly into the 2D Convolutional Neural Network (CNN), which does not extract depth detail features and just has a low recognition accuracy for the spectrograms with little differences. To address the issues, we propose an attention pooling for the recognition. It uses multiple one-dimensional (1D) convolution layers to accumulate data in a certain dimension (time or frequency) of 2D spectrograms. Then we get the 1D feature accumulation maps and train them with 1D CNN classifier. The experiments show that the 1D CNN with attention pooling based channel accumulation achieved 90.88% average accuracy, while the traditional 2D CNN is 78.97%. The former is 11.91% higher recognition accuracy than the latter, which can be applied to harmonic radar target classification.