Research And Implementation On Plant Image Segementation Algorithm Based On Neural Network

In recent years,image segmentation technology has been widely used in security,medicine,agriculture and other fields as an important part of image analysis.Traditional segmentation technology has disadvantages such as relying on manual experience and unable to achieve fully automatic image segmentation.With the development of deep learning,neural network based image segmentation not only greatly improves the accuracy,but also achieves fully automatic end-to-end segmentation.The research in this paper is based on the species identification project on the grassland.The main body of the plant to be identified is segmented from the image by segmentation technology,and the interference of non-target areas is eliminated to provide the basis for plant species identification and plant growth status analysis.The actual growth environment background of plants is often more complicated,and the parts used for main identification are different.At present,relevant algorithms lack generalization for plant image segmentation.In this paper,the plant image segmentation is mainly divided into leaf-type plants and flower-type plants to establish their respective data sets.Through model comparison,the following improvements are needed for the U-Net model with better performance based on actual projects: In view of the problem that the number of network model layers is small and the feature information extraction is not comprehensive,the feature fusion between each layer is enhanced by adding residual blocks in the contraction and expansion part of the U-Net network,and at the same time,the over-fitting problem in the model training process is avoided;In addition,considering the multi-scale problem of plant images collected in the field,the multi-scale input mechanism is applied in the contraction part of the network so that the model can obtain more complete semantic information for both small-scale and large-scale images,and enhance the generalization ability of the model;In view of the problem that the traditional U-Net network receives the shallow plant feature information of the corresponding contraction channel in the model expansion network part,it will also merge the redundant feature information that is useless for the final segmentation target.A dual attention mechanism is added to the network,that is,the channel attention mechanism and the spatial attention mechanism,allows the network to learn more important plant regions for segmentation,and improve the efficiency and accuracy of model segmentation.Finally,through the segmentation effect verification of the improved multi-scale Res-Att-UNet model,it is proved that the method can achieve good segmentation results in both the flower-type plant data set and the leaftype plant data set.