Image Analysis Method For Complex Traits Of Rice Grain And Panicle Based On Deep Learning

Rice is one of the most important crops,and its yield is related to the survival and health of human beings.The cultivation of high-yield rice has always been an important direction of rice breeding,and the measurement of rice phenotypic traits is the key link in rice breeding.The rice phenotypic traits mainly include grain traits and panicle traits.The former includes grain number,grain area,grain length and grain width.The latter includes panicle length,primary branch length and grain number on panicle.The traditional methods of obtaining phenotypic traits mainly adopt artificial measurement.The traditional methods of acquiring phenotypic traits mainly adopt artificial measurement which has the disadvantages of large amount of labor,low efficiency and low measurement precision.Using the image processing technology,the measurement efficiency and accuracy can be improved and the automation can be realized easily.Based on image processing technology,we designed a phenotypic traits measurement platform which combines grain traits measurement and panicle traits measurement.On this platform,most of the parameters of grain traits and panicle traits can be measured accurately.For grain number on panicle,traditional image processing method is difficult to measure accurately because of complex adhesion and occlusion between grains of panicle.In order to solve this problem,a prediction method based on deep learning for grain number of panicle is proposed in this study.In this method,fully convolutional neural network and grain center labeling method were applied to detect and count grains on intact panicle without damaging the morphological structure of the panicle.The measurement error of grain number on panicle reaches 3.47%,which satisfies the measurement error index.For the covered grain,the traditional image processing method cannot measure their traits parameters.These grains are usually ignored,which affects the integrity of the measurement results.In this study,the method of grain traits reduction based on deep learning was proposed.By simulating the phenomenon of grains occlusion,an automatic annotation method is designed to generate training samples.A convolutional neural network,which has the function of image reduction,is trained to restore the fragmentary portion of the covered grain image.The reduced grain is very close to the normal grain in shape and color,and can be used to measure the grain length,grain width and area.Through the experiment,the difference between the grain traits parameters between the reduced image and the real image is compared,and the feasibility of this method in practical application is analyzed.The prediction method of grain number on panicle based on deep learning,which was proposed in this study,can overcome the problem of grain adhesion and occlusion.In view of the problem that the traits parameters of the covered grains can not be measured,the method of grain traits reduction based on deep learning provides an effective solution.