Improved image semantic segmentation algorithm based on EMA

Aiming at the lack of semantic correlation between the parameters of expectation maximization attention(EMA) algorithm and images and the lack of attention to inter-channel information, a dual attention network EMA+ algorithm was proposed. Two modules were designed: spatial attention module and channel attention module. The EMA algorithm was used as the main structure by the spatial attention module. In the responsibility estimation step, the feature map itself was used as the initial parameter in the expectation maximization(EM) algorithm, and the semantic association between the parameter and the feature map was increased. Efficient channel attention(ECA) was used in the channel attention module by using one-dimensional convolution to learn the interactive information between channels. It avoided breaking the direct correspondence between channels and their weights due to dimensionality reduction operations. EMA+ significantly improved semantic segmentation tasks’ performance by fusing spatial attention modules and channel attention modules. The experimental results showed that EMA+ has achieved better intersection-over-union than EMANet and other methods on PASCAL VOC 2012 and some more complex datasets, and had better generalization ability.