SAR Image Target Recognition Based on Improved Hybrid Attention

As for the problem of low SAR image target recognition rate, the present paper improved the single module of residual attention network. Firstly, max pooling, average pooling and stochastic pooling were combined to put forward a dynamic hybrid pooling method for inaccurate middle and down sampling of mask branch to make the weight of hybrid attention of mask branch extraction more accurate; and then, channel attention mechanism was added to the trunk branch to enhance the weight of the useful feature, so as to improve the efficiency of information flow. The experiment based on MSTAR data set indicated that, compared with other algorithms, the improved model was relatively accurate. The accuracy was 1.16% higher, at the same time, the size of the improved model was only 1/3 of residual attention network.