基于类原型网络及动态语义注意力的小样本语义边缘检测OA

Accurate semantic edge information is crucial for machine vision applications such as object reconstruction,image generation,and medical imaging.However,existing deep learning-based semantic edge detection methods require a large amount of annotated data to train the network model,and the cost of obtaining a sufficient number of labeled samples is prohibitively high.To overcome this limitation,this paper proposes a few-shot semantic edge detection method based on class-prototype network and dynamic semantic attention,named DSAPNet(dynamic semantic attention and prototype network).It employs a meta-learning training strategy to locate semantic edges of novel categories using only a few annotated samples.The proposed network incorporates a class-prototype-based small-scale semantic segmenter,which provides multi-scale semantic information to the semantic edge detector at different stages of the network architecture.A dynamic semantic attention module is designed,which generates multi-scale spatial attention maps to highlight semantically meaningful regions and focuses the decoder module on those regions,effectively supplementing missing semantic information at the edges.A receptive field enhancement block and a context information fusion module are constructed,improving the receptive field of the output features in the encoding path and integrating their contextual information.Experiments on two few-shot semantic edge detection datasets,i.e.,FSE-1000 and SBD-5i,reveal the effectiveness of the proposed DSAPNet in the N-way K-shot few-shot semantic edge detection task.