Optimization of Category Unbalanced Semantic Segmentation Based on Geometric Structure Loss
Category imbalance has been a challenging problem in semantic segmentation tasks.To solve such problems,an op-timization method based on geometric structure loss is proposed in this paper.The geometric structure loss function is proposed for the problem of poor segmentation of few-pixel categories.This loss function first extracts the geometric structure information between categories using the labeled graph information of the model,where the sum of the number of pixels of contours of the same class is used as the contour perimeter of the class,and the sum of the number of pixels of adjacent contours of neighboring classes is used as the adjacency area between categories.Then,the geometric structure information is converted into a directed graph structure by a dy-namic balanced normalization operation with the contour perimeter as the node information of the graph and the adjacency area as the connection relationship of the nodes.Finally,the discrepancy between the predicted labeled and real labeled directed graphs is calculated to obtain the specific loss function values,which are used to optimize the model for parameter learning.The geometric structure loss function proposed in this paper can dynamically adjust the contribution of pixel less categories to the overall loss and better utilize the contextual structure information,thus effectively improving the overall segmentation performance of the model.
国家科技期刊平台
NSTL
万方数据
