Crop Image Detection Based On Deep Learning

In recent years,with the rapid development of agricultural science and technology,the total grain output has increased day by day,but the problem of crop disaster prevention and control has become increasingly serious.Fast and accurate crop image detection technology is of great significance for crop diseases detection,and crop growth period monitoring.In the field of crop disease detection,traditional methods mainly rely on directly using artificial feature operators such as Histogram of Oriented Gradient(HOG)and Local Binary Patterns(LBP).These method do not have high detection accuracy when the disease image texture features are more complex.At present,detection methods based on deep learning have better detection performance through multiple training on images.Therefore,a complete image training database needs to be established.However,in actual situations,many crop disease images are sparse and difficult to collect,and there are few publicly available standard crop disease image databases for use.One way to improve deep learning methods that rely too much on image sample data is to use GAN to generate sample images.However,the commonly used generative adversarial networks result in poor image quality.Limited by the number and diversity of samples,there is still much room for improvement in detection accuracy.In the field of crop growth period detection,taking rice as an example,remote sensing satellite image data is currently mainly used to detect the growth period of rice.This method is easily affected by the weather and the system is huge.In addition,due to the complex texture features of the image during the growing period of rice,and the small differences in the details of different images,the deep learning method based on convolutional neural networks also encountered a bottleneck in the use process.This thesis has carried out research on the above problems.The related research work and innovation are as follows:(1)Established a database of diseased crops mask images and rice growing period image databases and uploaded them to the Internet for future generations to use,the database link:https://github.com/Kamiko Liu.(2)Image detection of crop diseases based on improved artificial feature operators.The Local Line Vector Pattern(LLVP)operator is proposed by fusing different kinds of improved artificial features and applying them to the task of crop disease image detection in complex environments,which improves the detection accuracy under traditional methods.(3)By introducing texture and light mask,the network structure and loss function are improved,and a light mask generative adversarial networks(LMGAN)is proposed to generate high-quality crop disease images with artificially controllable shape,texture and illumination brightness.According to the experimental results,it is verified that with the help of LMGAN,the accuracy of crop image detection methods based on deep learning is improved.(4)Image detection of rice growing period based on improved convolutional neural network.Firstly,using a convolutional layer to replace the fully connected layer,making the network's requirements for input images more flexible,and realized the calculation of the target rice area.Then introduced a flexible connection layer based on Unet,and optimized the loss function of the network.Finally,a crop convolutional neural network(Crop-Net)is proposed,which improved the detection accuracy.Experiments show that the proposed method can improve the detection accuracy of crop images.However,in the field of deep learning,network structures are diverse.In the experiments of this thesis,it is found that the detection performance of the network is easily affected by the image quality.Therefore,in the future,we will try to further optimize the network structure,introduce noise modules,and improve the anti-interference ability of the network.