Study on Loss Functions for Semantic Segmentation Based on Imbalanced Datasets
In the task of image semantic segmentation,the class imbalance distribution of the dataset is a common phenomenon;Constructing a more effective loss function to mitigate its adverse effects is of great research value.This paper takes the loss function as the research object,improves the scale measure of the Focal loss function using the spatial distribution characteristics and correlation of the training sample pixels,and proposes an Adapt-Focal loss function with adaptive adjustment of weights based on iterative batches.Quantitative experiments were completed using three classical convolutional neural networks,DeepLabV3+,U-Net and RWL-ENet,with the class imbalance SAR image dataset GDUT-Nansha as the experimental object.The experimental results show that the proposed Adapt-Focal loss function effectively improves the segmentation accuracy IoU and PA values for the few-sample feature class bare ground and roads compared with the CE,FL,dfl,acw,lovász and softiou loss functions;meanwhile,the o-verall segmentation accuracy metrics mPA and mIoU are both substantially improved.The effectiveness of Adapt-Fo-cal loss function in semantic segmentation of category imbalanced images is verified.
SAR imageSemantic segmentationClass-imbalancedLoss functionAdaptive adjustment of weights
NETL
NSTL
万方数据
