Multi-Scale Localization Grouping Weighted Weakly Supervised Video Instance Segmentation and Air Cruiser Application

Implementing video instance segmentation (VIS) to detect, segment, and track targets based on vision system is important research for air cruiser. Large data with high sampling difficulty result in inefficient network training and limit the air cruisers in adapting to natural scenes during mission....

Full description

Saved in:
Bibliographic Details
Main Authors: Yunnan Deng, Yaomin Liu, Yinhui Zhang, Zifen He
Format: Article
Language:English
Published: MDPI AG 2025-04-01
Series:Applied Sciences
Subjects:
Online Access:https://www.mdpi.com/2076-3417/15/7/4025
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Implementing video instance segmentation (VIS) to detect, segment, and track targets based on vision system is important research for air cruiser. Large data with high sampling difficulty result in inefficient network training and limit the air cruisers in adapting to natural scenes during mission. A multi-scale localization grouping weighted weakly supervised VIS (MLGW-VIS) is proposed. Firstly, a spatial information refinement module is designed to supplement the multi-scale spatial location information of the high-level features of the feature pyramid. Secondly, feature interaction among the channels in each sub-space of mask features is strengthened by grouping weighting module. Thirdly, projection and color similarity loss are introduced to achieve weak supervised learning. The experimental results on the data from YouTube-VIS 2019 show that MLGW-VIS has increased the average segmentation accuracy by 5.7% and reached 37.9%, and has achieved positive effects on the perception and location accuracy of objects on the air cruiser platform.
ISSN:2076-3417