Research On Wheat Counting Method Based On Multi-scale Residual U-net Network And Attention Self-adversarial Network

Wheat counting has always been an important part of the regional yield estimation and wheat phenotypic analysis in China and the world.Because the accurate prediction of wheat production in the early stage mainly relies on agricultural workers to carry out manual counting in the field,this counting method is not only time-consuming,but the data has very strong subjectivity and statistical methods lack a uniform standard.These are not conducive to the country’s accurate estimation and analysis of wheat production,and then affect the country’s guidance and decision-making on wheat production.Therefore,automatic wheat counting has attracted the attention of researchers,and automatic wheat counting is a major challenge in computer vision.Effective wheat counting methods need to solve problems such as complex backgrounds,severe target adhesion,mutual occlusion,and uneven illumination changes in wheat images.Most of the traditional wheat counting methods are based on the manual extraction of some feature information,such as color feature,shape feature,gradient histogram feature and so on.Although these feature information can roughly identify the location of the wheat ear,the recognition effect of the model will be worse in the case of complex background and illumination change.With the rise of deep convolutional neural networks(DCNN)and its powerful feature extraction capabilities,it has been studied by a large number of agricultural researchers and applied to all aspects of agriculture.In order to better achieve wheat counting,this thesis studies the deep convolutional neural network,and builds a multi-scale residual deep convolutional neural network for wheat counting tasks based on this,which improves the accuracy of wheat counting.At the same time,the wheat ear image in natural scenes has the problem of mutual occlusion.The deep convolutional generative adversarial networks(DCGAN),because of its unique adversarial mechanism,allows the network to learn the occluded targets,thus solving the occlusion problem of wheat.In this thesis,a deep convolutional attention self-adversarial network is constructed to solve the problem of occlusion of wheat ears,and a good counting accuracy is obtained on the open wheat dataset.The main research results of this thesis are as follows:1.A wheat counting method based on multi-scale residual U-net network is proposed.In order to improve the accuracy of wheat counting,the network first uses a residual module and recurrent convolution blocks to construct multi-scale residual blocks to obtain richer context information to identify wheat ears.Second,in order to have strong learning ability,the classic encoding and decoding structure of u-net network is used to integrate the features under the multi-scale and increase the receptive field of the network.Finally,the proposed multi-scale residual block is fused into this codec structure,which further enriches the network’s ability to learn features.It is verified by experiments that the proposed multi-scale residual U-net network has a certain improvement in wheat counting accuracy on the self-built wheat dataset compared to the U-net network.2.A wheat counting method based on attention self-adversarial network is proposed.In order to further solve the problems of occlusion and self-similarity among wheat targets,and further improve the accuracy of wheat counting.The network first adds an attention model behind a single hourglass network to form an attention hourglass network.This network structure pays more attention to the phenotype of wheat and weakens the attention to areas outside wheat;the same attention hourglass network makes it adversarial network,and thus builds a complete attention self-adversarial network.Finally,it is verified by experiments that the proposed network structure has improved accuracy in wheat counting compared with other network structures.At the same time,the composition analysis of the network is also performed,which verifies that the introduction of the attention mechanism and the addition of the adversarial network are both effective.