Research On Soybean Flower-Pod Recognition And Falling Flower And Pod Law Based On Deep Learning

The flowering of flowers and pods is one of the main factors affecting soybean yield,but its genetic mechanism is still not clear.Rapid,accurate,and quantitative investigation of flower and pod phenotypes is the key link to elucidate this mechanism.At present,the phenotype investigation of falling flowers and pods mainly relies on the gauze laying-gauze method or interval artificial observation method,which has problems such as poor real-time performance,strong subjectivity,and low flux.The existence of these problems makes it urgent for us to acquire flower pod phenotypic information with high throughput,high accuracy,and high automation.At the same time,with the rapid development of deep learning technology,the design and development of deep networks suitable for crop phenotype acquisition have become a new research hotspot in the field of modern intelligent agriculture.In this paper,a new deep network model was designed to count the flowers and pods of soybean automatically,and then the law of falling flowers and pods of soybean was preliminarily explored.The specific work is as follows:This paper proposes a deep learning method for visually detecting soybean flowers(Soy Flower RCNN).According to the field environment of flowers and their own characteristics,the Faster R-CNN algorithm was improved,and the original feature extraction network VGG16 was replaced by Res Net-50.Experimental results show that the accuracy of the Soy Flower RCNN model proposed under the m AP evaluation index of 0.5 thresholds is 2.2% higher than that of the benchmark model,and the detection accuracy of other target detection algorithms(SSD,YOLOV3)is also higher than that of other target detection algorithms.The average accuracy of 94.5% was obtained on the test set(rain,illumination,multi-scale,etc.).The square of the Pearson correlation coefficient is 0.809 when comparing the predicted value of the model with the real value of manual counting.Pod counting mainly faces two challenges.On the one hand,pod texture and color are similar to a leaf,main stem,and weed,leading to confusion between theme and background recognition,which increases the difficulty of detection;on the other hand,pod shapes and sizes are different at different stages of growth and development,which increases the difficulty of model recognition and detection.To address these challenges,based on Faster R-CNN,this paper improved by using Res Net-50 as the basic feature extraction network to build a CSPRes Net-50 composite network to extract high-level semantic features of pods and add a feature pyramid network to the structure to enhance the detection ability of pods of different sizes.Propose the Soypod RCNN model.KMeans algorithm was used to cluster pod length and width,and the preset anchor box more in line with pod shape was fitted.The results show that the Soy Pod RCNN model has good detection ability for pods at different growth stages.The average accuracy is 91.1% in the verification set under various complex scenes.The square of the Pearson correlation coefficient is 0.9046 in the comparison between the predicted value of the model and the real value.Based on Soy Flower RCNN and Soy Pod RCNN models,this paper proposes a fusion model that can count flower pods at the same time.This model has the best performance compared with other deep learning networks.Then,based on the fusion model,the identification,location,and counting of flowers and pods of the tested varieties(HN51 and DN252)during the whole growth period were carried out.The results showed that the flowering and flowering rates of different varieties were different,and the slow-flowering varieties were more productive,which was consistent with the results of previous studies.Secondly,the spatial distribution of flowers and pods of different varieties was analyzed,and the phenomenon of falling flowers and pods was found at each node.From the bottom to the upper region,the number of falling flowers and pods decreased successively,and the rate of forming pods increased successively.Finally,the proportion of falling flowers and falling pods in different growth ranges were analyzed,and it was concluded that there were phenomena of falling flowers and falling pods in each period.The proportion of falling flowers increased gradually with the passage of time,and the proportion of falling pods increased first and then decreased with the change of time.Based on the deep learning target detection model design of the algorithm can accomplish effectively replace artificial flower counting tasks.The use of the model algorithm will greatly promote the research on the laws of the fallen petal fall pod basic genetic law and its resolution,and by way of marker-assisted selection breeding high yield varieties with the low rate of fallen petal fall pod important phenotypic technology base.