基于轻量级U型网络的遥感影像分割方法研究

Research on Remote Sensing Image Segmentation Method Based on Lightweight U-shaped Network

全卷积神经网络的出现,让遥感影像分割技术不断进步,但是这些方法常常因为两个问题备受限制。首先,由于编解码网络在跳跃连接部分往往会引起特征冗余,导致深度网络模型无法学习到有用的特征信息;其次,主流的深度网络模型致力于提高遥感影像的分割精度,因此通常采用更为复杂的编码策略导致网络模型参数量巨大,资源消耗多。为解决以上问题,论文提出了两个网络优化策略,首先在网络的跳跃连接部分添加注意力机制模块,使网络能够学习到更加有用的知识,从而提升网络的特征学习能力;其次使用深度可分离卷积代替常规卷积以减少网络参数量,在保证网络对遥感影像分割精度的同时提升网络的泛化能力。基于上述策略,该文章设计了一种新的用于遥感影像分割的轻量化网络。实验结果表明,使用提出的网络模型对遥感影像进行图像分割,准确率达到82.92%,MIoU达到69.85%,模型参数量仅有9.835MB,计算量26.028GFlops。

The emergence of full convolutional neural network has made continuous progress in remote sensing image segmentation technology,but these methods are often limited by two problems. First of all,the deep network model cannot learn useful feature information due to feature redundancy in jumping connections. Secondly,mainstream deep network models are committed to improving the segmentation accuracy of remote sensing images,so more complex coding strategies are usually adopted,resulting in a large number of network model parameters and high resource consumption. In order to solve the above problems,this paper proposes two network optimization strategies. Firstly,the attention mechanism module is added in the jump connection part of the network,so that the network can learn more useful knowledge,so as to improve the feature learning ability of the network. Secondly,deep separable convolution is used instead of conventional convolution to reduce the number of network parameters and improve the generalization ability of the network while ensuring the segmentation accuracy of remote sensing images. Based on the above strategy,a new lightweight network for remote sensing image segmentation is designed. Experimental results show that the proposed network model can achieve 82.92% accuracy and 69.85% MIoU,with only 9.835MB of model parameters and 26.028GFlops of calculation.