Research On Weed Detection Algorithm Based On Convolutional Neural Network

Agriculture plays a very important role in a productive life and is the primary condition for all production activities.In the process of crop production activities,weeds are the biggest threat.Weeds are not only vigorous,but also compete with crops for nutrition,sunshine,and so on.With the constant update and progress of science and technology,concepts such as precision agriculture and smart agriculture have been proposed.Therefore,it is very important to study accurate and rapid weed detection methods.However,there are still some deficiencies in the existing research on weed detection,such as large amount of network model parameters,missing and false detection of small-scale crops and weeds,which make the detection effect of the existing weed detection algorithms not ideal.Therefore,based on the theory of image processing and deep learning,this thesis takes sugar beet and its associated weeds as the research object.The main work can be summarized as follows:(1)Aiming at the problems of large parameters of the SSD network model,poor detection effect of small objects,and low detection accuracy of crops and weeds,a weed detection algorithm based on multi-scale fusion module and feature enhancement is proposed in this thesis.Firstly,in SSD model,a lightweight network MobileNet is used to extract the image features,replacing the original feature extraction network VGG16 to reduce the scale of model parameters;then a multi-scale fusion module is designed to improve the small target detection effect.In other words,the key information in the shallow feature map is enhanced through the channel attention mechanism,and then the receptive field of the feature map is expanded by the expansion convolution of different expansion rates,so as to fully obtain the context information of the target,and finally,the two branches are fused to make the shallow feature map contain not only more small target detail information,but also rich semantic information when detecting small targets,so as to solve the problem of missed detection of small targets;on this basis,the channel attention mechanism is used to enhance the features of the six output feature maps.The experimental results show that the weed detection model based on multi-scale fusion module and feature enhancement proposed in this thesis can achieve an average detection accuracy of 88.84%in the beet and weed image dataset in the natural environment,which is 3.23%higher than the standard SSD model,57.09%fewer parameters,and 88.44%faster detection speed.At the same time,the detection ability of the model for small target crops,weeds,and leaf overlap is improved.(2)In view of the huge amount of parameters and computation of the convolutional neural network model,which makes it difficult to apply the convolutional neural network to real-world scenarios,a weed detection algorithm based on dynamic pruning neural network is proposed in this thesis.Firstly,the importance of each channel in the network is evaluated with the help of SE module,and sparse regularization is applied;then an adaptive penalty weight of network sparsity is proposed.According to the learning effect of the current model,the weight is dynamically adjusted and added to the final training target to realize the dynamic compression of the model.Finally,the proposed model compression method is verified by experiments,and the experiments are worked out on the classic multi-classification dataset CIFAR-10.From the experimental results,it can be concluded that the weed detection algorithm based on dynamic pruning neural network proposed in this thesis fully exploits the redundancy of the network.The amount of parameters is reduced by 43.97%,the amount of computation is reduced by 82.94%,and the loss of accuracy is small.Subsequently,the proposed model compression method is applied to the improved model proposed in(1),and experiments are carried out on the sugar beet and weed datasets.It is proved that the method can also have a good effect on target detection tasks,and the detection accuracy of the model after pruning is basically unchanged.It can also be concluded that a lightweight network model like MobileNet can also be further compressed,and the compressed network model can still maintain a high detection accuracy.