Research On Fine-grained Image Recognition Of Zanthoxylum Rust Based On Multi-scale Image Segmentation

With the continuous development of computer vision,the agricultural application of artificial intelligence technology is increasing.Zanthoxylum is widely planted in China and it is a high-value-added economic forest product.Zanthoxylum rust is the main factor in affecting the yield and quality of Zanthoxylum.It has important practical value to identify the infection degree of Zanthoxylum.Due to the complex color and shape of the Zanthoxylum rust,the accuracy of judging the severity of rust which depends on the traditional methods of forestry experts is not high.Aiming at the above problem,this thesis studies the method of segmentation and recognition of Zanthoxylum rust based on deep learning.First,the Goog Le Net network is used to judge whether the leaves have rust,then we study the fine-grained segmentation method of Zanthoxylum rust leaves.This thesis proposed a model with higher segmentation accuracy based on the Deep Lab V2 model.The main contributions of this thesis include the following:1.Images of Zanthoxylum rust were collected and pre-processed by data screening.The images of Zanthoxylum rust were labeled according to leaves,convex spores and brownish spots.Leaf images were divided into 4 categories including background.In this thesis,100 leaves of Zanthoxylum rust were labeled with finegrained features.After cropping,the size of each image is 480 × 360,and the total number is 5753.This thesis constructs a high-quality image dataset of Zanthoxylum rust to provide data support for the training of the deep-learning model of Zanthoxylum rust segmentation.2.In view of the small area and complex shape of most of the Zanthoxylum rust,this thesis proposed the FASPP(Five-branch Atrous Spatial Pyramid Pooling)module,and proposed a segmentation model of Zanthoxylum rust based on FASPP module.The FASPP module improves the recognition effect of small targets in the rusty area by adding a branch with smaller atrous rate to fully extract the multi-scale information of the image.We compare the performance of the model on different parameters of the new branch.Based on the fine-grained image segmentation dataset produced in this thesis,it is obtained by experiment that when the new branch has a atrous rate of 3,the effect is maximized.Compared with the experimental result of the original model,the recognition accuracy of the convex sporesis is increased by 2.68%,and the recognition accuracy of the brownish spots is increased by 3.96%.The MIo U is increased by 1.59%.3.Pixel value processing is performed on the original image segmentation dataset to produce a differential image dataset,and an image segmentation model with enhanced features of rust area is proposed.The differential image and the original image are input to the network as two branches of the network model.Compared with the experimental result of the model based on FASPP module,the recognition accuracy of the convex sporesis is increased by 1.16%,and the recognition accuracy of the brownish spots is increased by 3.24%.The MIo U is increased by 1.63%.4.Considering the efficiency problem in practical applications,this thesis appropriately reduces the number of convolution layers in branch network.Without degrading the segmentation performance of the model,the thesis made experimental simulation and proposed an image segmentation model with faster image processing speed.It can process 22 pictures of Zanthoxylum rust leaves in a minute.The processing speed has been improved to a certain extent.