| Since ancient times,China has been a major agricultural country focusing on agricultural production,and the problem of plant diseases and insect pests has been one of the main problems that have plagued efficient agricultural production.With the introduction of precision agriculture,in order to achieve high output,in addition to scientific sowing and efficient protection,fast and accurate methods for identifying plant diseases and insect pests have become an urgent need.At the same time,with the upgrading of computer hardware,the computing performance has become increasingly powerful,providing the possibility of building an automatic identification system on the computer.In addition,with the popularization and development of mobile devices,comprehensive factors such as high-definition cameras and high-performance processors provide an unprecedented scale of development for disease diagnosis based on automatic image recognition.Image recognition systems based on convolutional neural networks have achieved good classification results in computer vision,but relying on a large number of labeled training sets.In order to alleviate the above-mentioned problems,this thesis introduces an unsupervised generative confrontation network into the plant image recognition system.The main research contents are as follows:1)Constructing image recognition data set of plant diseases and insect pests.Under the premise of ensuring that the effective information of the image is unchanged,the data of 15 types of plant diseases and pest images are enhanced and the data set is expanded.At the same time,filtering and noise reduction enhance the effective components.2)Improvements in generative adversarial networks.In order to obtain better quality of generated samples and recognition effect,the GAN network and its derived models are improved respectively.For the GAN model,in order to obtain deeper image features,a convolutional neural network(CNN)is introduced to construct a deep convolutional generative adversarial networks(DCGAN);for the CGAN model,in order to extract effective information of image diversity,Gaussian mixed noise is introduced to replace Gaussian noise with single distribution;for the SSGAN model,in order to improve the discriminative ability of the discriminative network and better distinguish the generated sample category from the real data category,the classification result of the Softmax classification function is adjusted from K + 1 to 2K.For the ACGAN model,it can be regarded as the synthesis of the CGAN and SSGAN models.In order to extract more subtle deep image information,a two-way discriminant network is introduced.3)Based on the powerful unsupervised data enhancement capability of the deep convolution generation adversarial network and the in-depth study and inspiration of the field of plant disease image recognition,considering comprehensively the variants and improved models of generative adversarial networks,a new type of comprehensive improved generative adversarial networks(GMM-SMP-DAGAN)is proposed for data enhancement and image recognition.In order to improve the classification effect of the network,the idea of model fusion is introduced to build a high-performance image recognition system.Model training was carried out in Plant Village,a public plant data set after data cleaning and data enhancement,with an accuracy rate of 97.68% and 98.14%,respectively.In order to verify the reliability and generalization ability of the comprehensive improvement generation adversarial network GMM-SMP-DAGAN mentioned in this thesis,the transfer learning Alex Net network was used to set up a comparative experiment.Under the same experimental environment and data set,the GMM-SMP-DAGAN mentioned in this thesis has better performance than the transfer learning Alex Net network.Finally,transplanting the trained model to the mobile phone terminal,which can identify the plant diseases and pest images in real time and efficiently. |
PDF Full Text Request