智能系统学报2025，Vol.20Issue(3)：605-620,16.DOI:10.11992/tis.202405021
基于主干网络浅深层特征的无人机海上分割算法
Unmanned aerial vehicle-driven sea segmentation based on the shallow and deep features of the backbone
摘要
Abstract
To improve the real-time and accurate segmentation of targets during the autonomous landing of UAVs in complex marine environments,studying the impact of backbone and shallow/deep features on the performance of al-gorithms is crucial.Based on the DeepLabV3+framework,a Shallow and Deep Features of Backbone(SDFB)al-gorithm is established for maritime scene segmentation.First,to address the issue of reduced target stability caused by wind-wave disturbances,a feature extraction method is proposed by optimizing the MobileNetV2 structure,and this method resolves the issue of low processing speed of single frame images in the algorithm.Second,to address the issue of numerous deep feature output channels and the uneven distribution of atmospheric turbulence noise,a feature filter-ing mechanism is proposed by selectively aggregating features using local and global information,thereby eliminating redundant features while solving the high sensitivity issue of the algorithm to environmental noise.Third,to address the issue of uneven lighting reducing the clarity of target boundaries,a parallel contour learning mechanism is established by extracting contour information from shallow spatial dimensions and deep channel dimensions,thereby solving the low-efficiency issue regarding the utilization of contour features.Finally,to address the issue of background occlusion disrupting the integrity of target features,a multi-scale feature fusion mechanism is established through the fusion op-timization of strip pooling,and this solves the connection issue of the algorithm to discrete distribution features.Finally,relevant experiments reveal that the LMSC algorithm exhibits higher real-time accuracy than other algorithms and can better adapt to the segmentation requirements of UAVs in maritime scenes.关键词
复杂海上场景/语义分割/无人机降落/船舶目标/DeepLabV3+/注意力机制/深度学习/卷积神经网络Key words
complex maritime scene/semantic segmentation/drone landing/ship target/DeepLabV3+/attention mech-anism/deep learning/convolutional neural network分类
信息技术与安全科学引用本文复制引用
沈昊,葛泉波,吴高峰..基于主干网络浅深层特征的无人机海上分割算法[J].智能系统学报,2025,20(3):605-620,16.基金项目
国家自然科学基金项目(62033010,62303233) （62033010,62303233）
江苏省高校青蓝工程项目(R2023Q07). （R2023Q07）
