Forest image segmentation method based on improved DeepLabv3+
Forest resource management planning has gained increasing attention due to the vital role that forests play in ecological balance,climate regulation,and the provision of various resources.Due to the complex structure of forests and their fragmented distribution,it is difficult to accurately distinguish between tree areas and non-tree areas and it is difficult to accurately predict forest areas.To solve the problems of difficult extraction of forest area and inaccurate boundary segmentation,based on UAV remote sensing images of forest areas,an improved DeepLabv3+model was proposed to investigate the intelligent and accurate extraction of forest areas,to provide technical support for intelligent monitoring and management of forest area.Firstly,the backbone network Xception in the model was replaced with a lightweight MobileNetv2 to reduce the calculation of model parameters and improve the operation efficiency of the model.Secondly,the CFF(cross feature fusion)module was used to fuse the multi-scale low-level and high-level features of the backbone network and the cavity convolution in the encoder stage to obtain high-resolution mask fea-tures and effectively aggregate the multi-level encoder features.Thirdly,the cSE(spatial squeeze and channel excita-tion)channel attention module was introduced in the decoder stage by establishing the dependency relationship between channels.The cSE model can better obtain the features on different channels,emphasize the useful features,suppress the useless features,and improve the expressiveness of the network to pay attention to the edge position of the input image,so as to improve the accuracy of forest area segmentation and the efficiency of the model.Finally,the deep features after convolution were fused with the shallow features to enhance the segmentation performance of the network.The results showed that the average pixel accuracy of mPA model for forest categories obtained based on the improved DeepLabv3+deep learning neural network segmentation model reached 93.85％,the average intersection and merger ratio mIoU reached 89.17％,and the accuracy reached 95.66％,which were 0.77％,1.8％and 0.89％higher than the original DeepLabv3+network.The number of model parameters was reduced by 48.84 M,and the model detection speed FPS was increased by 17.93 frames per second,detection efficiency was enhanced and segmen-tation performance was improved.
万方数据
维普
