Research On Semantic Segmentation Of Forest Tree Image Based On Convolutional Neural Network

As an important natural resource of the country,forest resources play a key role in ecological protection such as climate regulation and water conservation.Once destroyed,forest resources can be very hard to restore while causing great distress to the maintenance of the ecosystem.How to accurately segment forest areas,monitor the dynamic changes of forest resources in real-time and manage forest resources intelligently and scientifically has been considered one of the research hotspots in forestry industry.With the rapid development of convolutional neural networks in computer vision,it also promotes to make its extension to the segmentation and extraction of forest images.With the purpose of further strengthening the ecological forestry construction and promoting the sustainable development of forestry economy.Applying convolutional neural networks to segmentation studies of forest images,and the main research contents are as follows:Traditional image segmentation methods have problems such as low accuracy,time consuming and difficult to deploy on large-scale.Convolutional neural networks have significant advantages in improving the accuracy of forest segmentation with their powerful autonomous learning ability and efficient feature fitting capability.The method of forest image semantic segmentation based on U-Net network is proposed.The model uses transposed convolution to recover the image resolution,and the unique U-shaped structure to combine the feature map of the encoder and the feature map obtained by up-sampling the corresponding decoder.To reconstruct the image features,skip connection bridges the low-level feature maps with the high-level feature maps,which effectively fuse the feature information of the forest trees.Through comparing and analyzing the traditional image segmentation method with the predicted segmentation map of forest trees obtained by the proposed U-Net network,the accuracy rate of U-Net is as high as 88.75%,which experimentally verifies the accuracy and effectiveness of the proposed network.Considering that the convolved forest feature map lacks a targeted feature information extraction process,it cannot be accurately segmented to confirm whether it belongs to the forest area for fine areas,fuzzy forest boundaries and shadow occlusions,which would easily result in wrong segmentation and missed segmentation situations.Based on U-Net,an improved semantic segmentation model of forest images with multi-scale fusion(Pyramid Feature Extraction-UNet,PFE-UNet)is proposed.A pyramid extraction module is designed in transition layer of network,which captures contextual information of multi-scale receptive fields with different dilation rates of dilation convolution to fuse multi-scale features.Moreover,attention modules i.e.channel-wise attention module and spatial attention module are designed to maintain different mapping relationships between channels and enhance spatial relationships between features,respectively,which highlight features for specific segmentation tasks while suppressing irrelevant regions.To further reduce the computational cost of the model,novel convolutional unit is proposed to replace the original standard convolutional block.The lowrank decomposition of the convolution kernel and changing the order of the convolutional layers guarantee the speed of iteration to some extent.The experimental results demonstrate that the attention mechanism,pyramidal feature extraction,and the introduction of asymmetric depthwise separable convolution provide the improved PFE-UNet model with significant advantages in dealing with small areas,discontinuous areas,and fuzzy boundaries.By comparing experiments with other convolutional neural networks in multiple groups,the accuracy of the improved PFE-UNet model is as high as 94.23%,which validates the superiority and advance of the improved PFE-UNet network,which provides a beneficial reference for scientific monitoring of dynamic changes in forest resources.