Semantic Segmentation Based on Dividing Different Difficulty Class Levels
-
摘要: 语义分割的任务是将给定图片的每一个像素进行分类,为了解决当类数目急剧增多或类的特征频繁改变的时候,语 义分割的准确性会急剧地下降的问题。本研究提出将像素分类任务按照分类难度划分成不同的子任务。具体工作分为两部分: 为每一个子任务训练一个神经网络,训练一个集成神经网络。根据图像像素的多少来划分难度等级的数量。通过为每一个不同 的难度等级训练神经网络,可以获得各个子任务的概率图,然后通过这些概率图来训练集成网络。在实验部分,本研究将数据集 上的11个类别划分成容易、中等、困难三类进行训练,在CamVid数据集上使用平均IoU衡量该方法语义分割准确率。实验结果 表明,本研究方法与单一U-net对比传统方法对比在各类平均IoU上有了2%的提升。尤其是在围栏,人行道,自行车手这三类上 有超过5%的提升。Abstract: The task of semantic segmentation is to classify each pixel of a given picture, in order to solve the problem that the accuracy of semantic segmentation will drop sharply when the number of classes increases sharply or the characteristics of classes change frequently. This research proposes to divide the pixel classification task into different subtasks according to the classification difficulty. The specific work is divided into two parts: training a neural network for each subtask and training an integrated neural network. Divide the number of difficulty levels according to the number of image pixels. By training the neural network for each different difficulty level, the probability map of each subtask can be obtained, and then the integrated network can be trained through these probability maps. In the experimental part, this research divides the 11 categories on the data set into three categories: easy, medium, and difficult for training. The average IoU is used on the CamVid data set to measure the accuracy of semantic segmentation of this method. The experimental results show that this research method has a 2% improvement in the average IoU of various types compared with the single U-net compared with the traditional method. Especially in the three categories of fences, sidewalks, and cyclists, there is an increase of more than 5%.
-
Key words:
- Semantic segmentation /
- Sub-tasks /
- Integration classifiers
-
[1] 许必宵,宫婧,孙知信.基于卷积神经网络的目标检测模型综述[J].计算机技术与发展,2019,29(12):87-92. [2] Ren S,He K,Girshick R,et al.Faster R-CNN:Towards RealTime Object Detection with Region Proposal Networks[J].IEEE Transactions on Pattern Analysis & Machine Intelligence,2017,39(6):1137-1149. [3] Purkait P,Zhao C,Zach C.SPP-Net:Deep Absolute Pose Regression with Synthetic Views[C]//British Machine Vision Conference(BMVC 2018).2017. [4] 罗会兰,陈鸿坤.基于深度学习的目标检测研究综述[J].电子学报,2020,48(6):1230-1239. [5] 赵永强,饶元,董世鹏,等.深度学习目标检测方法综述[J].中国图象图形学报,2020,25(4):629-654. [6] 刘军,后士浩,张凯,等.基于增强 Tiny YOLOV3 算法的车辆实时检测与跟踪[J].农业工程学报,2019,35(8):118-125. [7] 刘博,王胜正,赵建森,等.基于 Darknet 网络和 YOLOv3 算法的船舶跟踪识别[J].计算机应用,2019,39(6):1663-1668. [8] 戴伟聪,金龙旭,李国宁,等.遥感图像中飞机的改进 YOLOv3 实时检测算法[J].光电工程,2018,45(12):84-92. [9] 赵文清,严海,邵绪强.改进的非极大值抑制算法的目标检测[J].中国图象图形学报,2018,23(11):1676-1685. [10] Bodla N,Singh B,Chellappa R,et al.Soft-NMS:Improving Object Detection With One Line of Code[J].2017. [11] 侯志强,刘晓义,余旺盛,等.基于双阈值 - 非极大值抑制的Faster R-CNN 改进算法[J].光电工程,2019,46(12):82-92. -
点击查看大图
计量
- 文章访问数: 120
- HTML全文浏览量: 19
- PDF下载量: 9
- 被引次数: 0
下载: