Wheat Head Recognition Based On Improved Target Detection

The automatic detection technology for wheat ear has high scientific research and application value in yield estimation,seed screening,density estimation and gene morphology expression,etc.The main research focus is on the wheat ear recognition technology.In the past,the utilization rate of deep learning in recognition research is not high,and the recognition in the past is more dependent on manual classification and machine learning methods,which brings many problems,such as low detection accuracy,high labor cost,long detection training time,etc.,and it is difficult to achieve good results in the actual scene application.After the introduction of deep learning convolutional neural network model,some of these problems have been solved,but the efficiency and detection time of the basic model are not particularly ideal,so in order to better adapt to the needs of practical application scenarios,the original model needs to be improved and studied to enhance the efficiency of the model.This study focused on the GWHD dataset,which was led by 9 institutions and 7 countries.The training dataset covered multiple regions around the world with different varieties,planting densities,styles and field conditions.In this study,the training prediction of the original model was carried out respectively for the two existing target detection methods called Faster R-CNN and YOLO.In addition,some improvements were made to the original model to obtain the improved model with better detection effect and improve its generalization performance,so that the model can be better applied to other scenarios.The main research contents of this thesis are as follows:(1)Based on the latest GWHD world wheat spike data set,a total of 3423 images were used to construct the basic data set required by the experiment.After that,the basic data set is enhanced and expanded,and the data set is expanded by rotation,inversion and clipping,and the expanded data set is used as the experimental data set for experiments.(2)According to the two major methods of target detection,Faster R-CNN and YOLO.The research is divided into two parts and the basic models in two general directions are used to extract features from the data set and train models,and the corresponding experimental results are obtained.Some important experimental evaluation indicators,such as m AP and Loss,were recorded during the experiment,which was convenient for comparative analysis of experimental results after the experiment.(3)Some improvements are made to the existing Faster R-CNN deep learning algorithm.ResNet50 network is used to replace VGG16 network in feature extraction network,and BiFPN weighted fusion unit is introduced in ResNet50 network.At the same time,Kmeans clustering algorithm was added to the model,and the aspect ratio of nine prior frames was obtained by clustering the target frames to replace the fixed aspect ratio of prior frames in the original algorithm,so as to better adapt to training.Finally,the improved experimental results are compared with the original model,verify the feasibility of improvement.(4)the existing YOLO deep learning algorithm is improved to some extent.CSPdarknet is used to replace the original network in the feature extraction network,Kmeans clustering algorithm is added to the prior frame anchor,prior frame parameters suitable for the experimental data set are adjusted,and Mosaic data enhancement and cosine annealing learning rate algorithms are also added.Finally,the improved experimental results are compared with the original model,verify the feasibility of improvement.(5)The improved model is compared and analyzed with the current popular models in the field,and the superiority of the improved model for the data set in this thesis,namely the small target data set,is proved through the comparison of m AP and detection speed.