Research On Semantic Segmentation System Under Long-tail Data Distribution

With the development in the field of deep learning,there is a growing demand for semantic segmentation models in terms of tail category accuracy,such as in the fields of autonomous driving,medical image segmentation and so on.Although the research of long-tail recognition has achieved good results in classification,the development in the field of semantic segmentation is very limited,and the traditional long-tail recognition solutions are not suitable for solving the semantic segmentation task.This paper aims to study and solve the long-tail recognition problem in the semantic segmentation and build a semantic segmentation system which can improve the recognition of tail categories while satisfying the detection performance of each category.The main works in this paper are as follows:1.For the problem that the number of tail category data is seriously insufficient,this paper designs the Tail Data Augmentation(TDA)method,which can expand the number of pixels in the tail category without increasing the data volume.The experimental results show that the TDA scheme designed in this paper can improve the accuracy of the overall and especially the tail category,and the average accuracy of the tail category(mAccTail))can be improved by 1.28 percentage points on the basis of FCN-R50.2.For the current problem of low accuracy of the tail category of semantic segmentation network,this paper designs a semantic segmentation network SegCCRW(Segmentation with Classifier Calibration and Re-weighting).The network adaptively corrects the classifier and designs a reweighting loss to improve the learning weight of the tail category.Experiments show that the SegCCRW model improves the overall accuracy of the dataset,especially the tail category,and mAccTail can be improved by nearly 3 percentage points based on FCNR50,and can be further improved by about 4 percentage points after incorporating the TDA strategy.3.This paper designs the system functions and interface according to the actual application requirements,and builds a semantic segmentation system that can improve the accuracy of tail category.The system has good performance on tail category,and is suitable for scenarios with high requirements on tail category detection effect.