Abstract: The object edge details are difficult to retain due to the confusion of labels and the loss of small objects in traditional clothing image segmentation. To solve the issue, a semantic segmentation network model termed residual expectation maximization attention based on convolution attention feature is proposed. The proposed model first uses ResNeXt-50 as the backbone network for shared features, and introduces a set of parallel convolutional attention modules in the feature extraction stage, which can effectively suppress invalid features and make the features of the target region more prominent. Then, the residual idea is used to optimize the expectation maximization attention unit to avoid the gradient explosion or disappearance in the iterative process, so as to establish the relationship between the positions in the feature map and realize the semantic segmentation model based on saliency fusion learning. Finally, qualitative and quantitative experiments verify the efficiency of the proposed model on the traditional national costume data set. The mean intersection over union segmentation index reaches 83.91%, which achieves the best results among similar algorithms.

Key words: deep learning, traditional costume, feature extraction, attention mechanism, semantic segmentation