Anchor-Free Object Detection Network via Spatial Perception and Multiple Attention Weighted Fusion
To solve several problems such as limited spatial perception ability and rough feature fusion strategy in the existing U-shaped object detection framework,an Anchor-Free object detection network based on spatial perception and multiple attention weighted fusion is proposed.Firstly,the primary feature extraction module is utilized to obtain low-level details of more representa-tion areas.In the next step,the global residual perception module with multi-level feedback structure is designed to extract multi-level context from high-level semantic feature,which can significantly improve the multi-scale spatial perception of the con-textual areas.Then,the rough features are selectively processed by multiple attention iterative fusion module,which aims to guide the efficient feature fusion of low-level features and high-level semantic features,helping the network obtain both high-resolution and semantic features.Meanwhile,generalized intersection over union loss is added into the loss,teaching the network to predict re-sults in a more accurate and efficient way.Sufficient experiments are conducted on two datasets including standard MSCOCO 2017 and PASCAL VOC 2007.The results demonstrate the effectiveness of the proposed approach.
deep learningobject detectionAnchor-Free networkmultiple attention fusionweighted iterative fusion
国家科技期刊平台
NSTL
万方数据
