Multi-attention Guided Activation Propagation in CNNs

摘要

CNNs compute the activations of feature maps and propagate them through the networks. Activations carry various information with different impacts on the prediction, thus should be handled with different degrees. However, existing CNNs usually process them identically. Visual attention mechanism focuses on the selection of regions of interest and the control of information flow through the network. Therefore,we propose a multi-attention guided activation propagation approach (MAAP), which can be applied into existing CNNs to promote their performance. Attention maps are first computed based on the activations of feature maps, vary as the propagation goes deeper and focus on different regions of interest in the feature maps. Then multi-level attention is utilized to guide the activation propagation, giving CNNs the ability to adaptively highlight pivotal information and weaken uncorrected information. Experimental results on fine-grained image classification benchmark demonstrate that the applications of MAAP achieve better performance than state-of-the-art CNNs.