Research On Cherry Tomatoes Target Detection And Classification Based On Convolutional Neural Network

It is an inevitable trend of agricultural development to introduce mechanization and intellectualization in the process of agricultural production.As an important link in agricultural production,the sorting of fruits and vegetables mainly relies on visual inspection and comparison,which is characterized by high labor intensity and low efficiency,and cannot be unified due to the influence of individuals.In addition,if normal fruits and rotten fruits are piled together in the storage process,the deterioration rate of normal fruits will be accelerated to a certain extent.In order to prolong the preservation period of cherry tomatoes,the fruits need to be classified and stored as early as possible,and the sorting of fruits is best completed in the picking stage.At present,the automatic fruit grading method based on machine vision is generally carried out under ideal conditions,and the research object is placed in a simple background for detection,but when put into practice,the effect is often not up to expectation.At the same time,with the continuous pursuit of the performance of convolutional neural network,the number of layers of the network is also deepening,and the cost of computing equipment is also increasing,which is more detrimental to the promotion of agricultural mechanization and intelligence.In this thesis,cherry tomatoes is taken as the research object,deep learning technology is applied to automatic sorting of cherry tomatoes,and the powerful feature extraction ability of convolutional neural network is used to complete target detection and classification of cherry tomatoes.When making the data set of relevant experiments,the images involved in the real picking process will be simulated as much as possible.In order to reduce the cost of computing equipment,this thesis uses Shufflenet and Mobile Net as the backbone network of the model,and carries out a series of optimization measures to further speed up the training and detection speed of the model,and sets up the relative comparative test.Mainly completed the following work:(1)Production of data sets: obtain as many cherry tomatoes images that meet the experimental requirements as possible through Internet downloading and self-shooting,etc.Based on this,two data sets are made for target detection experiment and classification experiment respectively.For the data set of target detection experiment,Labelimg software is used to label cherry tomatoes in the image one by one.According to the corresponding standards,the data sets of classification experiments are divided into two categories: good and bad.(2)Constructing cherry tomatoes target detection algorithm:In this thesis,cherry tomatoes target detection model is built on the basis of Yolov3.Shufflenet is used to replace Darknet-53 as the trunk network of Yolov3,and the number of anchor boxes is reduced according to the actual demand.K-means clustering is used to get the prior boxes of different scales in6.GIOU algorithm is introduced as a measure to distinguish positive and negative samples.In order to accelerate the training speed of the model,the method of multi-process and GPU calculation is used.(3)Constructing cherry tomatoes classification algorithm: This thesis optimizes the network structure of Mobile Net V1 based on the fast down-sampling strategy.In the shallow layer of the network,the convolution with step size of 2 is continuously used to rapidly reduce the size of the feature graph and increase the number of channels,which not only reduces the amount of calculation,but also ensures the capacity of information.In order to reduce the training time,transfer learning method is used.And the generalization ability of the model is increased by means of data enhancement.A comparative experiment was set up to verify the effectiveness of the optimization measures.The model after training was tested.(4)Analysis of experimental results: the missed detection rate and repetition rate of the optimized target detection model were reduced by 2.4% and 1.7%,respectively.Especially when the fruit was partially shaded by leaves,the optimized model had better detection effect.The accuracy of the optimized cherry tomatoes classification model reached 96%,an increase of 4 percentage points.The optimized Mobile Net is 12.65% faster and takes less than one tenth of the time compared to traditional VGG-16 network detection.In this thesis,the convolutional neural network is used to complete the target detection and classification task of cherry tomatoes images.The production of data sets is more consistent with the actual application scenarios.The optimized model not only has a certain improvement in accuracy,but also has a significant improvement in training and detection speed.Therefore,this model can be better applied to mobile devices and embedded devices,and the calculation cost is relatively low when applied to actual production,which is conducive to the promotion and development of intelligent agricultural mechanization.