一种新型的无人机牧草分割网络—LMS-DeeplabV3+

A novel UAV pasture segmentation network—LMS-DeeplabV3+

目前草原环境复杂、牧草分散且与背景颜色差异小,无法实现高效精准的分割,因此本文提出了一种新型的轻量化多尺度DeeplabV3+网络(lightweight and multi-scale DeeplabV3+network,LMS-DeeplabV3+)。该网络以DeeplabV3+为基础网络,首先选用轻量级的MobilenetV2作为骨干网络用于初步特征提取,并为了适应牧草分割任务做了网络配置上的调整;其次在加强特征提取模块和解码模块中均使用深度可分离卷积代替普通卷积以轻量化网络;此外利用密集空洞空间金字塔池化(dense atrous spatial pyramid pooling,DASPP)模块捕获更大的感受野,加强各特征之间的交互;又引入卷积注意力机制(convolutional block attention module,CBAM)重分配权重加强特征提取。实验证明,提出的新网络与原始网络相比平均交并比(mean intersection over union,mIOU)提升了8.06个百分点、平均像素精度(mean pixel accuracy,mPA)提升了6.75个百分点,网络计算量和参数量均下降了90%以上,分割预测速度也有所提升,与其他主流分割网络相比各性能都表现更好。

Currently grassland environment is complex,pastures are scattered and have little difference in color from the background.It isn′t achieved the efficient and accurate segmentation.Therefore,this paper proposes a novel lightweight and multi-scale DeeplabV3+network(LMS-DeeplabV3+).The network uses DeeplabV3+as the base network,and first selects the lightweight MobilenetV2 as the backbone network for initial feature extraction and adjust its configuration to suit the pasture segmentation task;secondly the depth separable convolution is used instead of normal convolution in both the enhanced feature extraction and decoding modules to lighten the network;in addition,the dense atrous spatial pyramid pooling(DASPP)module is used to capture a larger sensory field and enhance the interaction among features;the convolutional block attention module(CBAM)is also introduced to reassign weights to enhance feature extraction.Experiments show that the proposed new network improves mean intersection over union(mIOU)by 8.06 percentage points and mean pixel accuracy(mPA)by 6.75 percentage points compared with the original network,reduces both the computation and the number of parameters of network by more than 90%,improves the segmentation prediction speed,and performes better in all aspects compared with other mainstream segmentation networks.