Phenotypic Parameters Acquisition Of Arabidopsis Leaves And Pods Based On Deep Learning

Crop phenotype studies reveal the effect of genome and environment interactions on plant traits.By analyzing the traits of crops,understanding the relationship between genotypes and phenotypes,selecting and cultivating crop varieties with excellent traits is the primary task of crop breeding.With the rapid development of image processing and artificial intelligence technology,related research on phenotypic parameter measurement based on crop image data has gradually attracted researchers' attention in the field of high-throughput phenotyping research.Image processing technology can avoid the damage caused to crops with manual measurement and the effects of subjective factors can also be avoided with the technology.As a model plant,Arabidopsis thaliana has a short cycle,which plays an important role in the research of increasing grain production and crop stress resistance.Therefore,in this paper,the model plant Arabidopsis thaliana was taken as the research object,and image processing technology was used to analyze Arabidopsis thaliana seedling leaf,Arabidopsis thaliana mature pod and other image data to realize Arabidopsis leaf segmentation and Arabidopsis pod detection.The relevant phenotypic parameters are also measured on this basis.The research contents carried out in this paper include:(1)Arabidopsis leaf segmentation.A data enhancement method based on image stitching and synthesis is proposed,which uses Mask R-CNN algorithm to complete single-leaf segmentation;and compared with traditional data enhancement methods to verify the effectiveness of this method;(2)Arabidopsis pod detection.In this experiment,the Faster R-CNN algorithm was used to realize the pod detection.According to the shortcoming of the non-maximum suppression algorithm and the characteristics of the Arabidopsis pod data,the improved non-maximum suppression and improved rotation proposals pod detection methods were proposed respectively;the F1 score of those methods reached 0.796 ?0.867 and 0.882,which was proved that both algorithms can detect overlapping cross pod,improve the pod detection recall rate,At the same time,the rotating box pod detection algorithm can better mark area of the pod,verify the effectiveness of this method;(3)Parameter acquisition.?Leaf counting.The correlation between the number of leaves after leaf segmentation and the results of manual statistics reached 0.877;?pod counting.The number of pod is an important factor in Arabidopsis yield,and its quantitative statistics can be also used to evaluate plant growth performance and reproductive adaptability,which has important biological significance.The correlation coefficient of the number of pods detected by artificial markers and detected by Faster R-CNN algorithm with improved anchor box size,improved non-maximum suppression algorithm and improved rotated proposals algorithm proposed in this paper reached 0.9147,0.9120,0.9661 respectively.? In addition,the automatic measurement of the leaf center point,leaf length,leaf width,and leaf area in the image space has been completed in the leaves divided,and the correlation coefficients between the calculated leaf length and leaf width and the manual measurement are all over 0.96;The detected pod has completed the automatic measurement of pod length,pod width and pod area in the image space,and the correlation coefficient between the calculated pod length and the artificial value is 0.9379.It can be known from the comparison and analysis of the measured values of the phenotype parameters and the artificial values that the measured values of the automatic parameter extraction algorithm proposed in this experiment have achieved good timeliness under the premise of using the previous segmentation and detection results.It has been proved that for the images of Arabidopsis leaves and pods,the relevant algorithms in this paper can be used to achieve leaf segmentation and pod detection non-destructively and accurately,and further completing automatic measurement of phenotypic parameters.Those can provide important methodological basis and data support for crop breeders,and provide methods and large-scale data support for phenotypic studies of rapeseed.