基于轻量级改进网络的无人机航拍图像语义分割研究

Research on semantic segmentation of aerial images based on improved lightweight networks

目前图像语义分割深度学习算法多为通用任务型,这使其应用于无人机航拍图像语义分割时存在目标尺度多变以及物体边界分割不清晰等问题。同时,与无人机相关的应用需要使用尽可能轻量化的网络。有鉴于此,提出了一种轻量化语义分割网络Lite-SFNet。采用轻量化的STDC2网络作为骨干网络,设计了一种轻量化的空间金字塔池化模块。通过减少金字塔分支数和引入高效的有效通道注意力(Efficient Channel Attention,ECA)模块,降低了模型参数量和提高了模型特征提取能力,进而提高了网络对多尺度目标的识别能力。对语义流校准模块(Flow Alignment Module,FAM)进行了改进,构建轻量化的解码器提高了网络对物体边界分割能力。在Aeroscapes等航拍图像数据集进行了仿真实验。实验结果表明:与现有轻量级模型相比,所提方法以较少的浮点计算量和参数量实现了较高的精度。

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.