Lightweight semantic segmentation algorithm based on multi-scale pooling and feature fusion
Semantic segmentation is an important component of visual understanding systems,which can recognize the content and location present in images.However,existing semantic segmentation algorithms face the challenge of balancing complexity and segmentation accuracy,and cannot be flexibly applied to practical scenarios.To address this issue,this paper proposes a efficient semantic segmentation algorithm based on multi-scale pooling and feature fusion,taking into account both performance and network parameters.The method uses Deeplabv3+as the main algorithm and an improved lightweight MobileNetV2 as the backbone network to reduce the complexity of the network model.Using a Distinctive Atrous Spatial Pyramid Pooling Module(DASPP),utilizing multi-scale pooling operations and atrous convolution operations of different sizes,fully capturing multi-scale target features and rich global contextual semantic information.In the decoding section,attention mechanism is introduced to enhance representation,and a Multi-level Feature Fusion Network(MFFN)is proposed to effectively fuse high-level and low-level features,further improving segmentation accuracy.The model proposed in this article greatly reduces the number of model parameters and significantly improves performance compared to classical semantic segmentation methods.Experiments are conducted on the PASCAL VOC 2012 dataset,and the number of model parameters in this paper is only 6.66 M,achieving an accuracy of 73.72%on the test set.
万方数据
维普
