Research On Crop Detection Algorithm Based On Convolution Neural Network

Crop detection is a key technology in intelligent agriculture,which plays a key role in automatic harvesting and determines whether the robot can quickly and accurately identify the specific location of crops.In the traditional crop detection,people mostly detect based on manually selected features such as color and texture of crops.However,these methods have low detection accuracy under complex features.In recent years,convolution neural network is outstanding in the field of target detection and can be used to detect crops.However,in the actual scenario,there are still many challenges in this technology.First of all,enough image data sets need to be made for network training before the experiment,but the data sets of most crops are difficult to collect,and this is also a very time-consuming task;secondly,the individuals of most crops are not only very small and densely distributed,and they all have complex detection backgrounds,which bring difficulties to the detection of crops.Finally,for the detection network,the shallow network computing reasoning is fast,but the target feature extraction is not sufficient.The deep feature extraction network is beneficial to the target feature extraction,but it will greatly increase the network parameters and the amount of calculation,and the training time is long.In view of this,this paper focuses on the model based on convolution neural network,taking kiwifruit and corn as an example,three works are done respectively,as follows:(1)Two kinds of data sets are made.In this paper,we made a total of 1100 kiwifruit data sets and 900 corn data sets,each picture contains multiple targets.This paper uses the annotation tool Label Img to annotate these two kinds of data sets.(2)A crop detection method based on transfer learning and Faster R-CNN is proposed.In order to solve the problems of low accuracy and long training time in the detection of two-stage Faster R-CNN,this paper first uses different classification networks as the backbone feature extraction network of Faster R-CNN,pre-trains the model with VOC2007 data set,and then uses the idea of transfer learning to load the model weight of pre-training to train the data set of this paper,and gradually fine-tune the network weight value.finally,the proposed method is compared with different algorithms.The results show that the method is feasible and effective.(3)A lightweight YOLO＿v4 algorithm is designed to improve the detection speed of two kinds of crops,and multi-scale information fusion is used to ensure the detection accuracy.First of all,in order to solve the problem of large number of network parameters caused by the large number of backbone network layers in YOLO＿v4,lightweight network Mobile Net＿v3 is used for feature extraction,and then deep separable convolution is introduced to calculate the parameters,which further reduces the size of network parameters.Finally,multi-scale information is integrated to enhance the characteristics of crops,improve the utilization of semantic information and ensure the accuracy.The experimental results show that the detection accuracy of the improved YOLO＿v4 algorithm is about 48 m less than that of the original algorithm,and the accuracy is also improved,which can better meet the needs of practical applications.To sum up,based on the convolution neural network method,this paper has done some research on the detection of corn and kiwifruit,and through a series of experiments to verify the effectiveness of this method,but also shows that the method has a certain practicability.