Research on semantic segmentation of aerial images based on improved lightweight networks
At present,deep learning algorithms for image semantic segmentation are mostly general task oriented,while aerial image semantic segmentation has problems such as variable target scales and unclear object boundary segmentation.In addition,drone related applications require the network to be as lightweight as possible.Therefore,a lightweight semantic segmentation network Lite SFNet is proposed,which uses the lightweight STDC2 network as the backbone network,and designs a lightweight spatial pyramid pooling module.By reducing the number of pyramid branches and introducing efficient ECA attention modules,it reduces the amount of model parameters and improves the ability to extract model features,and improves the network's ability to recognize multi-scale targets.The FAM module is improved,and a lightweight decoder is constructed to enhance the network's ability to segment object boundaries.Simulation experiments are conducted on Aeroscape and other aerial image datasets.Experimental results show that compared with the existing lightweight model,the proposed method achieves higher accuracy with less floating-point computation and parameters.
semantic segmentationlightweight neural networkssemantic segmentation of aerial imagesfeature pyramidfeature fusion