基于可变形卷积技术的街景图像语义分割算法
A Semantic Segmentation Algorithm for Street View Images Based on Deformable Convolution Technique
岳明齐 1张迎春 1吴立杰 1秦晓海1
作者信息
- 1. 北京工商大学计算机学院,北京 100048
- 折叠
摘要
目前图像语义分割算法中可能会出现分割图像的不连续与细尺度目标丢失的缺陷,故提出可变形卷积融合增强图像的语义分割算法.算法集 HRNet 网络框架、Xception Module 以及可变形的卷积于一体,用轻量级 Xception Module 优化HRNet原先存在的Bottleneck模块,同时在网络的第一阶段串联融合可变形卷积,通过建立轻量级融合加强网络从而增强针对细尺度目标特征物的辨识精度,从而使得该轻量级融合增强网络在粗尺度目标物被分割时取得相对多的细尺度目标的语义特征信息,进一步缓解语义分割图像的不连续与细尺度的目标丢失.使用Cityscapes数据集,实验结果可以说明,优化后的算法对于细尺度目标分割精度得到了显著的增强,同时解决了图像语义分割导致的分割不连续的问题.然后进行实验使用的是公开数据集PASCAL VOC 2012,实验进一步的验证了优化算法的鲁棒性以及泛化能力.
Abstract
The current image semantic segmentation algorithms may have the defects of discontinuity of segmented images and loss of fine-scale targets,so we proposed a deformable convolutional fusion enhanced image semantic seg-mentation algorithm.The algorithm integrates the HRNet network framework,Xception Module and deformable convo-lution,optimizes the existing Bottleneck module of HRNet with lightweight Xception Module,and fuses deformable convolution in the first stage of the network to enhance the recognition accuracy of fine-scale target features by build-ing a lightweight fusion enhancement network.This lightweight fusion-enhanced network obtains relatively more se-mantic feature information of the fine-scale target when the coarse-scale target is segmented,which further alleviates the discontinuity of the semantic segmented image and the loss of the fine-scale target.Using the Cityscapes dataset,the experimental results can illustrate that the optimized algorithm has significantly enhanced the segmentation accura-cy for fine-scale targets,while solving the problem of segmentation discontinuity caused by semantic segmentation of images.The experiments were conducted using the publicly available dataset PASCAL VOC 2012,which further vali-dates the robustness and generalization ability of the optimized algorithm.
关键词
图像语义分割/高分辨率网络/可变形卷积Key words
Image semantic segmentation/High resolution network/Deformable convolution引用本文复制引用
出版年
2024
NETL
NSTL
万方数据