英文摘要:
Aiming at the problem of low edge fitting accuracy in traditional cultural pattern segmentation models, this paper optimizes the segmentation model from the perspectives of model prediction and data labeling. Firstly, propose an iterative upsampling strategy based on edge prediction. In the stage of generating the prediction map, the pre-trained point classifier is fused with the shallow features of the network, and the pixels of the blurred edge are further classified, so as to obtain a better prediction map with high edge quality. Secondly, for the problem of label ambiguity or error in pixel labeling, we propose a hybrid loss function based on label relaxation, which is combined with the cross-entropy loss function to support the training process of the segmentation model. Finally, on the traditional cultural pattern data set, we simulate to verify the effectiveness of the algorithm proposed in this paper, and also proves that the algorithm has a strong fault-tolerant mechanism, and can better improve the segmentation quality.
|