边缘引导选择与特征聚合的巡飞器航拍动态目标实例分割OA
Edge selection guidance and feature aggregation for dynamic ob-ject instance segmentation in aerial photography of patrol aircraft
针对动态地物目标与背景特征类间相似以及弱边缘感知不足导致分割精度低的问题,构建边缘先验显式选择、空间-通道动态互补聚合,双域下采样高频补偿的特征建模链路,提出了边缘引导选择与特征聚合的巡飞器航拍动态目标实例分割方法.首先,设计多尺度边缘引导选择感知模块,将多尺度特征与局部均值之间的残差响应作为边缘先验并生成边缘响应图,再通过双域选择机制强化模型对动态目标弱边缘信息的感知能力.其次,构建自适应特征聚合模块利用通道混洗机制实现通道级交互以激活跨组语义信息,并聚合动态卷积与深度动态卷积自适应提取的空间语义信息与通道互补细节特征,以提高特征响应度从而解决目标类间特征相似的问题.最后,为克服普通卷积提取多尺度空间信息的局限性,设计双域下采样自适应模块融合双路径提取到的层级特征信息,实现多尺度信息的自适应聚合与特征强化.实验结果表明,本文提出的边缘引导选择与特征聚合的巡飞器航拍动态目标实例分割模型mAP50-95和mAP50分割精度分别达到51.8％和84.3％,相较于基准模型提高5.1％和4.6％,将模型部署在Jetson Nano B01移动开发平台,能够实现动态地物目标的分割.
Aimed at the problem of low segmentation accuracy caused by the similarity between dynamic ground targets and background feature classes,as well as insufficient weak edge perception,a feature mod-eling link was constructed,which included edge prior explicit selection,spatial channel dynamic comple-mentary aggregation,and dual domain downsampling high-frequency compensation.An edge guided selec-tion and feature aggregation method for dynamic object instance segmentation in aerial photography of pa-trol aircraft was proposed.Firstly,a multi-scale edge guided selection perception module was designed,which took the residual response between multi-scale features and local means as the edge prior and gener-ated an edge response map.Then,a dual domain selection mechanism was applied to enhance the model's ability to perceive weak edge information of dynamic targets.Secondly,an adaptive feature aggregation module was constructed to utilize channel level interaction through a channel shuffling mechanism to acti-vate cross-group semantic information,and to aggregate spatial semantic information extracted adaptively by dynamic convolution and deep dynamic convolution with complementary channel detail features,in or-der to improve feature responsiveness and solve the problem of feature similarity between target classes.Finally,to overcome the limitations of ordinary convolution in extracting multi-scale spatial information,a dual domain downsampling adaptive module was designed to fuse the hierarchical feature information ex-tracted from dual paths,achieving adaptive aggregation and feature enhancement capabilities of multi-scale information.The experimental results show that the edge guided selection and feature aggregation pro-posed in this paper for the dynamic object instance segmentation models mAP50-95 and mAP50 in aerial photography achieve segmentation accuracies of 51.8％and 84.3％,respectively,which are 5.1％and 4.6％higher than the benchmark model.Deploying the models on the Jetson Nano B01 mobile develop-ment platform can achieve dynamic object segmentation.
宫冕;张印辉;何自芬;陈光晨
昆明理工大学 机电工程学院,云南 昆明 650500昆明理工大学 机电工程学院,云南 昆明 650500昆明理工大学 机电工程学院,云南 昆明 650500昆明理工大学 机电工程学院,云南 昆明 650500
信息技术与安全科学
实例分割动态目标边缘引导选择自适应特征聚合模型部署
instance segmentationdynamic objectedge guided selectionadaptive feature aggrega-tionmodel deployment
《光学精密工程》 2026 (16)
2536-2558,23
国家自然科学基金(No.62561032,No.62171206)昆明理工大学分析测试基金(No.2024P20231103003)
评论