Research On Wheat Detection And Accurate Segmentation Based On Deep Learning

Wheat is one of the most widely grown staple foods in China and is of great significance to economic development and social stability.The statistics of wheat ears and geometric phenotypic measurements provide data support for wheat yield estimation,quality assessment and variety selection.The traditional statistics of the number of wheat ears depend on manual sampling,which requires a lot of manpower and material resources and a long cycle.In recent years,the improvement of the accuracy of image acquisition equipment has inspired researchers to use image processing methods to analyze and analyze wheat images during grain filling,but this method is not accurate and has poor applicability.The geometric phenotype of wheat ears needs to be obtained on the premise of precise segmentation.The measurement of spikelets needs to further achieve non-adhesion segmentation.There is still no effective solution in the current research.With the development of deep learning,the problems in the agricultural field can be solved.The solution is based on agricultural image data,through the introduction of target detection and image segmentation.Based on field wheat images builds and trains neural network models.It consists of two parts:one is the wheat ear detection model,which is used to predict the relative position of wheat ears in the image and the number of wheat ears;The second is the wheat-segmentation model,which realizes two-level segmentation of wheat ears per plant and obtains the segmentation results of wheat ears and spikelets.In order to improve the detection precision of wheat ears,this paper proposes an IoU cascade method,which can combine the results of Cascade R-CNN and Faster R-CNN,and use IoU threshold to increase the detection rate.Then the error check box is filtered through the LBP texture analysis method to reduce the false positive rate.On the wheat ear segmentation,the wheat segmentation and spikelet non-adhesion segmentation are realized by the example segmentation method Mask R-CNN,and the geometric phenotypic parameters of wheat ears are calculated by using Mask coordinates.This paper studies and discusses wheat ear detection,wheat ear precision segmentation,wheat ear detection and precision segmentation system implementation.The main contents are as follows:(1)We propose a wheat ear target detection algorithm based on IoU cascade,which enables the detection and counting of wheat ears in a single image.The existing algorithm has a low detection rate and a high false detection rate.To solve this problem,this study proposes a target detection algorithm.We use Cascade R-CNN as the basic detection algorithm,cascade the detection results of Faster R-CNN using IoU threshold,and then analyze the LBP texture of the detected wheat ears.In the experimental field,1079 wheat images were collected,and a total of 56182 two-dimensional minimum bounding boxes of wheat ears are manually labeled to construct a data set.Compared with the Faster R-CNN and Cascade R-CNN detection methods,the value of the method(AP-Average-Precision)reach 0.81,and the correct rate of single-image wheat ear count reach 92.8%,which is better than the existing methods.(2)A semantic segmentation model is constructed to achieve wheat segmentation and geometric phenotypic measurements.We construct an instance segmentation algorithm Mask R-CNN,including Deep Residual Network(ResNet),Region Proposal Networks(RPN),and Fully Convolutional Networks(FCN).Images are collected from 800 complete wheat ears samples and manually labeled,and data augmentation is performed as a data set.The AP value on this data set reach 0.848,while the composite index F1(F1-Measure)value is 0.830.Compared with the two baseline methods of OTSU and FCN,the method is better than the above two methods from the accuracy and P-R curve analysis.(3)We design a system that implements the detection and precise segmentation of wheat ears.For agronomists and researchers,we developed a prediction system for wheat ears detection and segmentation to measure wheat ears and geometric phenotypes.PyQt is used to construct a human-computer interaction interface,and the prediction model used is selected according to the image input by the user,and then the prediction result and information are displayed in the output window.