A multiple resolution branch attention neural network for scene understanding of intelligent autonomous platform

文献类型： 外文期刊

第一作者： Dai, Yingpeng

作者： Dai, Yingpeng;Meng, Lingfeng;Ren, Jie;Wang, Yutan

作者机构：

关键词： Information exchange; Lightweight neural network; Multiple resolution branch attention; Semantic segmentation; Unmanned platform

期刊名称：EXPERT SYSTEMS WITH APPLICATIONS （ 影响因子：7.5； 五年影响因子：7.8 ）

ISSN： 0957-4174

年卷期： 2025 年 267 卷

页码：

收录情况： SCI

摘要： Scene understanding is a key technology for autonomous platform to understand environmental information. For intelligent autonomous platform, it is required to find effective information in complex environments. Convolutional neural networks lack information exchange and selection between different features, which limits their ability to extract effective feature information. To tackle this problem, a lightweight multiple resolution branch attention neural network (MRBANet) is proposed for real-time vision tasks in complex scenes. First, a lightweight three branch network structure is designed to extract rich feature information. From the top branch to the bottom branch, channel width changes from narrow to wide and feature map resolution changes from high to low, so as to quickly extract features at different levels. The second is to exchange feature information among different branches. Here, the context feature information with different resolutions is extracted to enhance the feature representation of different branches. Furthermore, in the process of information exchange, the multiple branch attention module is designed to adaptively adjust the proportion of each branch fusion information to find salient context feature information. On the Cityscapes dataset, we achieve 75.8 % Mean Intersection over Union (MIoU) with 16.1 M parameters at the speed of 26.3 FPS on a single 1070Ti card. On the BDD, UAS, and Jujubes dataset, MRBAnet achieves excellent performance. On the real world, MRBANet is successfully deployed on the intelligent autonomous platform and achieves better semantic segmentation and classification performance.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序