Development Of Image Recognition System Of Lingwu Long Jujube

As a fresh fruit,Lingwu long jujube has short picking time and heavy workload.The migration of labor to the city and the aging of the population have made manual picking extremely difficult.Intelligent agricultural and forestry picking equipment is urgently needed to replace manual labor to improve picking efficiency.The intelligent recognition of Lingwu long jujube is the key to intelligent picking,and the accuracy of image recognition is the premise of the intelligent recognition system.Based on the practical problems such as the lack of robustness,low recognition accuracy and inaccurate target positioning of Lingwu long jujube in natural state,this paper designed a set of Lingwu long jujube image recognition system,which can efficiently and accurately identify nature The long jujube target under the conditions provides theory and data reference for image recognition in the research of forestry industry harvesting intelligent equipment.The main research methods and conclusions adopted are as follows:1.Establish Lingwu long jujube dataset.Under natural conditions,the Lingwu long jujube,which is in the mature season,is randomly sampled and sampled as a data set.The data set is the basis of deep learning and affects the accuracy of network training.A total of 27,081 pre-processed images were obtained.2.The double loss function improvement based on faster region convolutional neural network(Faster R-CNN).According to the migration learning results of the classic Faster R-CNN,it is found that the identified long jujube target cannot obtain satisfactory accuracy and processing time.Therefore,an improved single-loss Faster R-CNN network was first proposed,through a smaller dimension of network depth and The number of convolutions identifies the target and obtains higher real-time performance,but the accuracy of the recognition is not high.To this end,the algorithm is improved again,and the double-layer loss function is used in the Region Proposal Network(RPN)network.All eigenvalues obtain the corresponding probability distribution after experiencing the Softmax loss function,and then use it to limit the angle and other factors.The stronger A-softmax loss and L-softmax loss functions are performed in parallel to calculate the feature with larger probability value,and then map it to the feature image to obtain the Region of Interest(ROI),and finally the specific identification Out of the target area.The recall rate,accuracy rate and average accuracy of the obtained dual loss function Faster R-CNN network are 0.9462,0.9296 and 0.8,respectively,which are significantly improved compared to other networks.3.Image recognition system design.Based on the above recognition results and algorithms,a recognition system was developed.The system consists of two parts:hardware and software.The hardware uses an industrial computer as the control and calculation core to control the industrial camera to capture images under natural conditions,and combine GPU(Graphics Processing Unit)calculations to achieve real-time image processing.The software development uses the previously trained dual loss function Faster R-CNN network as the image detector,and designs a simple and easy-to-operate GUI(Graphical User Interface),which can display the scene images collected by the current camera and the results of real-time image processing in real time.Through the test,the average processing time of operating the image recognition system is 4.68 seconds.Counting the results of image processing by the system,the evaluation standard values obtained are:the recall rate is 0.9472,the precision is 0.9262,and the average precision is 0.8.Compared with the result of the double loss function Faster R-CNN,the error is within a thousandths.The image recognition system designed in this paper realizes the recognition of Lingwu long jujube under natural conditions,which lays the foundation for the image recognition system of picking robots.