Research On Generating Data Enhancement Technology Of Networks In Plant Disease Classification

The diseases of the fruit trees on the farm bring a lot of losses every year.If many infectious diseases can be detected and dealt with in an early stage,the losses on the farm can be greatly reduced,and the spread of the diseases can even prevent huge economic losses.With the continuous development of artificial intelligence technology,farm disease detection has also changed from manual patrols in the past,and the mode of expert detection has gradually developed into the current mode of relying on cameras to take pictures and algorithm recognition for disease detection.This can significantly reduce the loss caused by the unfamiliarity of the fruit growers with the disease.At present,various recognition algorithms are widely used in the recognition of plant lesions.Among them,deep learning algorithms have recently shown good effects in plant lesion recognition.However,due to quarantine restrictions,some highly contagious plant diseases will be burned as soon as the corresponding plants are discovered,which makes it difficult to get a great quantity of lesion data,which seriously affects the application of deep learning in the field of plant lesion identification.At the same time,due to different types of plant diseases,their lesions may be very similar in structure and other information,which makes it difficult to recognize deep learning algorithms.At this time,the leaf information must be combined to accurately identify the corresponding disease.Therefore,this thesis proposes to use GAN to generate plant lesion images to enhance deep learning network training.However,when GAN generates a complete plant diseased leaf image,because the complete image has too many details,it will generate an image with poor visual and algorithm evaluation results.Through experiments,it is found that the use of GAN can generate local lesions with better evaluation results,but the local lesions are directly combined with leaves and the evaluation results are poor.Therefore,this thesis focuses on the problem that GAN can only generate better local lesions,but deep learning requires complete lesion leaf training.Two methods for segmenting lesions after generating local lesions are proposed,and the completeness can be obtained in large quantities.The healthy plant leaf image is synthesized to obtain the complete diseased leaf image needed in this thesis to enhance the deep learning network training.The main research work of this thesis is as follows:(1)Add improved Deeplab v3 image segmentation network after improved WGAN-GP generates the generated image to solve the problem of how to generate local lesions and segment them.(2)Propose a Binary Generative Adversarial Network to solve the problem of how to generate local lesions and segment them.(3)Use edge smoothing and image pyramid algorithm to synthesize the segmented lesions and complete healthy leaves to obtain the complete diseased leaf image of the plant required in this thesis.(4)By using the synthesized complete plant diseased leaf image to enhance the training of Alexnet,the recognition accuracy of the classification network is improved,and the effectiveness of the algorithm in this thesis is confirmed.