摘要
实例分割在图像分类的基础上为每一个物体生成像素级别的分割掩码,是当前计算机视觉领域热门研究课题,也是极具挑战性的任务之一。针对当前算法存在的分割精度和鲁棒性不高等问题,提出了一种改进的SOLOV2算法。首先,以FCN(Fully Convolutional Networks for Semantic Segmentation)算法为整体框架,采用ResNext作为骨干网络,在不增加网络参数量和计算量的前提下可以有效提升网络的精度;其次,采用改进的NAS-FPN(Neural Architecture Search Feature Pyramid Network)作为特征金字塔网络结构,这是一种可以在FPN中进行特征图的搜索和组合结构,使网络可以重新搜索并融合已经提取的特征图,以此来解决网络不能充分感知特征图从而导致网络精度不高的问题;最后,通过调整超参数得到整个分割网络模型。通过在COCO2017数据集上与BDD100K数据集上进行实验分析比较可知,改进的基于SOLOV2实例分割算法精度达到41.8%,在兼顾实时性的同时网络精度提升了2.1%。通过实验证明改进的算法可以适应多种交通场景,可以完成交通场景目标的检测与分割。
Instance segmentation generates pixel-level segmentation masks for each object based on image classification,which is currently one of the popular research topics and challenging tasks in computer vision.To address the problems of poor segmentation accuracy and robustness of current algorithms,we propose an improved SOLOV2 algorithm.Firstly,FCN(Fully Convolutional Networks for Semantic Segmentation)is used as the overall framework,and ResNext is adopted as the backbone network,which can effectively improve the accuracy of the network without raising the number of network parameters and computational effort.Secondly,a modified NAS-FPN(Neural Architecture Search Feature Pyramid Network)is used as the feature pyramid network structure,which is a structure that allows the search and combination of feature maps in the FPN,so that the network can re-search and fuse the already extracted feature maps,as a solution to the problem that the network cannot fully perceive the feature maps and thus the network accuracy is not high.Finally,the whole segmentation network model is obtained by adjusting the hyperparameters.The experimental analysis and comparison on the COCO2017 dataset and the BDD100K dataset shows that the improved SOLOV2 instance segmentation algorithm achieves 41.8%accuracy,which improves the network accuracy by 2.1%while taking into account the real-time performance.It is proved through experiments that the improved algorithm can adapt to a variety of traffic scenes and can complete the detection and segmentation of traffic scene targets.
作者
曾浩文
汪慧兰
赵侃
王桂丽
ZENG Hao-wen;WANG Hui-lan;ZHAO Kan;WANG Gui-li(School of Physics and Electronic Information,Anhui Normal University,Wuhu 241002,China)
出处
《计算机技术与发展》
2023年第9期45-51,共7页
Computer Technology and Development
基金
安徽省自然科学基金(1708085QF133)
安徽师范大学创新基金项目(2018XJJ100)
安徽省智能机器人信息融合与控制工程实验室资助(IFCIR2020004)。
