首页|一种用于道路场景分割的轻量级特征融合网络

一种用于道路场景分割的轻量级特征融合网络

扫码查看
道路场景的语义分割存在着实时性和准确性相冲突的矛盾,融合多层次特征和多尺度上下文信息可以提升分割模型的性能.但是复杂的特征融合将消耗较多的计算资源,并且现有的方法在分割过程中常常忽略位置信息,导致分割效果不理想.为了解决以上问题,使用了一种高效的轻量级特征融合网络(LFFNet)进行道路场景分割,具体来说就是使用了一个多层次特征融合模块,通过在注意机制中嵌入空间位置信息来增强多层次特征之间的语义一致性,以便在捕获远距离依赖关系的同时保留准确的位置信息.此外还使用了一种轻量语义金字塔模块,通过深度可分离卷积提取多尺度上下文信息.实验结果表明,LFFNet与现有的方法相比FLOPs减少了2.3倍,速度提高了1.7倍,在分割精度和计算效率上有较好的平衡.
A Lightweight Feature Fusion Network for Road Scene Segmentation
There is a contradiction between real-time and accuracy in the semantic segmentation of road scene.Integrating multi-level features and multi-scale context information can improve the performance of the segmentation model.However,complex feature fusion will consume a lot of computing resources,and the existing methods often ignore the location information during the segmentation process,which results in unsatisfactory segmentation performance.In order to solve the above problems,an efficient light feature fusion network(LFFNet)is used for road scene segmentation.Specifically,this paper uses a multi-level feature fusion module to enhance the semantic consistency by embedding spatial location information in the attention mechanism,so as to retain accurate location information while capturing long-range dependencies.Additionally,a light semantic pyramid module is utilized to extract multi-scale contextual information through depthwise separable convolutions.Experimental results demonstrate that LFFNet reduces FLOPs by 2.3 times and increases the speed by 1.7 times compared with existing methods and achieves a balance between segmentation accuracy and computational efficiency.

feature fusioncoordinate attentionmulti-scale context informationdeep separable convolutionsemantic seg-mentation

李富华、吴陈

展开 >

江苏科技大学计算机学院 镇江 212003

特征融合 坐标注意力 多尺度上下文信息 深度可分离卷积 语义分割

2024

计算机与数字工程
中国船舶重工集团公司第七0九研究所

计算机与数字工程

CSTPCD
影响因子:0.355
ISSN:1672-9722
年,卷(期):2024.52(8)
  • 1