Improvement of attentional mechanism for few-shot semantic segmentation

The rapid development of deep learning has brought significant improvements to semantic segmentation, but it usually requires a large amount of pixel-level labeled data to train the model. Therefore, the few-shot semantic segmentation algorithm based on meta-learning comes into being. By learning the segmentation model with generalization ability on the known class data, the model can accurately segment the unknown class images with a small amount of labeled samples. It meets the expectation of image segmentation to improve the generalization performance and reduce the data dependence of the model, and has a broad development prospect. This paper selects the self-support matching model (SSP) as the baseline model and proposes three improvements to the current algorithm: introducing the attention mechanism, optimizing the learning rate decay strategy and improving the loss function. In this paper, the proposed improved module and adjustment strategy are verified by several groups of experiments with Pascal VOC 2012 as data set. Compared with the original method, the accuracy of the improved few-shot semantic segmentation method has been improved by about 1.4%.