Feature extraction method and device based on second-order multi-scale attention mechanism

The invention provides a feature extraction method and device based on a second-order multi-scale attention mechanism. Through a parallel network structure, feature semantic information of two sub-networks is subjected to explicit construction of a dependency relationship between channels to increase the sensitivity of a model to information channels, and a second-order statistical value of pooling features is implicitly calculated, so that complex appearance and motion correlation which cannot be captured by global average pooling can be captured. Secondly, on the basis of network branches which reserve information in two aspects of channels and space to enhance cross-dimension interaction, multi-scale convolution kernels are adopted to effectively extract multi-scale space information with finer granularity, and meanwhile, a space dependency relationship with a longer distance can be established; finally, the two learned attention feature maps are multiplied and fused with the input primary feature map at t