Research On Semantic Segmentation For Green Crops Based On Deep Neural Network

The steady development of agricultural production is inseparable from the combination of science and technology.In the process of realizing smart agriculture,a more accurate segmentation of green crops is essential.The green crops segmentation models based on deep learning technology play a significant role for the realization of intelligent agricultural production.The existing green crops recognition mainly relies on human eye recognition and human experience to make judgments,which cannot guarantee real-time performance and the human resources are expensive.In view of these,this thesis aims to study the deep learning based semantic segmentation methods to accurately segment green crops and their flowers,which can be cooperated with agricultural robots to realize intelligent irrigation,fertilization and other agricultural production activities.It can be seen that the research in this thesis has certain theoretical significance and strong practical application value.In this thesis,we first propose a general semantic segmentation model and then apply such model to the green crops dataset to achieve green crops segmentation.The main research works and results in this thesis are listed as follows:(1)This thesis explores the relationship between atrous convolutions working in cascade and parallel modes,and deduces the equivalent relationship of atrous convolutions in these two connection modes.When the semantic segmentation models adopt the atrous convolutions with a too large dilation rate,the performance will drop sharply.In view of this,we propose SS-ASPP(Serial Sparse-Atrous Spatial Pyramid Pooling)module to replace ASPP(Atrous Spatial Pyramid Pooling)module in Deeplab V3+.Specifically,multiple atrous convolutions with different dilation rates in parallel mode are equivalently replaced with multiple atrous convolutions with same dilation rates in cascade mode,without reducing the receptive fields.Considering the object multiscale problem in semantic segmentation,some forward short connections are added between the atrous convolutions working in cascade mode to capture the multi-scale context information.The experimental results show that the proposed SS-ASPP module outperforms the ASPP module and the Dense ASPP module.On the Cityscapes dataset,the SS-ASPP module improves the mIoU(mean of class-wise Intersection over Union)index by 0.65%,compared to the ASPP module.(2)This thesis proposes a symmetrical encoder-decoder structure.We analyze the existing encoderdecoder structure,and observe that the decoder in Deeplab V3+ is relatively simple which results in the lack of detailed semantic information.In view of this,we improve the decoder to enrich the feature interaction between both encoder and decoder.In the new decoder,four lateral long connections from the encoder to the decoder are used,rather than one lateral long connection used in Deeplab V3+,which produces a symmetrical encoder-decoder structure like the U-Net model.The experimental results show that the proposed encoder-decoder structure certainly further improve the segmentation performance,and the mIoU index is improved by 0.48% compared with Deeplab V3+ on Cityscapes dataset.(3)This thesis proposes a new semantic segmentation model based on deep neural network.This model uses the proposed SS-ASPP module and symmetrical encoder-decoder structure,which not only can extract multi-scale context information but also is able to recover the detailed semantic information.We compare the proposed semantic segmentation model with the existing excellent semantic segmentation models on both Cityscapes and ADE20 k datasets.On the Cityscapes dataset,when Res Net50 is used as the backbone network,the mIoU score on the val set is 80.87%,with an improvement of 1.19% compared to Deeplab V3+;at the same time,the training time of the proposed semantic segmentation model is reduced by 11%-15% compared with Deeplab V3+under the same configuration.On the ADE20 k dataset,when Res Net50 is used as the backbone network,the mIoU score on the val set is 42.61%,with an improvement of 1.76% compared to Deeplab V3+.(4)This thesis applies the proposed semantic segmentation model to a green crop dataset to accomplish green crop segmentation.Since potato is one of the characteristic crops in Inner Mongolia,this thesis builds a green crop dataset for potatoes.The establishment of such dataset includes two parts: data collection and data labeling.Data collection is performed indoors and outdoors at different growth periods of potatoes and then Labelme software is used to manually label the segmentation masks of green crops and flowers.We train,verify,and test the proposed semantic segmentation model on this dataset.Experimental results show that the overall mIoU score on the val set is 90.07%,and the overall mIoU score on the test set is 90.55%,which verifies that the proposed model is certainly able to segment the green crops and their flowers.