Reserch On Wheat Powdery Mildew Spore Image Recognition And Segmentation Algorithm

Wheat powdery mildew is one of the common diseases in China,which is harmful to crops.It seriously affects the yield of wheat,causes the loss of economic property and affects food security.In order to realize detection and control of wheat powdery mildew early,and reduce its damage to crops and economy,it is necessary to identify the spores in the microscopic image,and then analyze the disease.Microspore image of wheat powdery mildew collected from the field,often with many other disease spores and dust debris,background interferences,traditional spore recognition technology for more interferences,debris which color and shape similar to the target image effect is not ideal,the target spore recognition rate is in low level,and more missed detection,false detection phenomenon in detection result.Traditional segmentation algorithms,such as threshold segmentation,K-means clustering,watershed algorithm,are easy to cause wrong segmentation,and which is sensitive to noise.The number of data sets of wheat powdery mildew spore image is small,and there are only more than 200 sample images that can be used for annotation training.Through investigation,there is little research on deep learning algorithm in spore recognition and segmentation,and most of the research objects are microspore images under Petri dish,which basically have no interference,The main research contents are as follows:1.To solve the problem of less effective samples in the data set of wheat powdery mildew spores,this paper uses the traditional data enhancement methods such as flipping,color transformation and so on to expand the data.At the same time,in order to increase the diversity of samples,the improved dcgan model is used to generate the image of wheat powdery mildew spores.Compared with the original model,the number of network layers is deepened and the residual structure is added,The wassertei distance was used to replace the loss function of the original model,and the network layer parameters were normalized,which improved the quality of the generated image and increased the amount of effective sample data.The improved model can effectively improve the recognition effect of the identification model for wheat powdery mildew spores.2.In order to solve the problem that the original SSD model has poor recognition effect for powdery mildew spores of small targets,and it is easy to miss target recognition for images with more interferences,this paper improves the SSD model,and proposes a multi-scale feature fusion module to fuse different scale feature maps,and expands the receptive field of shallow feature map combined with the designed receptive field module,The recognition rate of the model for wheat powdery mildew spores was effectively improved.3.In order to solve the problem that traditional segmentation methods are sensitive to noise,and there are serious over segmentation and wrong segmentation problems in the segmentation result,and it is difficult to deal with adhesion spore segmentation,this paper improves the U-Net model,uses a PPM-UNet network combined with pyramid pooling module for image segmentation,and the pyramid pooling module pools and fuses the deep feature maps in different scales,The improved model can effectively improve the segmentation accuracy of wheat powdery mildew spores,and combine with adaptive Canny edge detection algorithm to get a single image of wheat powdery mildew spores,which is convenient for subsequent counting work.