Study On Identification And Location Method Of Cassava Stem In Field

In order to solve the problem of pushing down and crushing cassava stalks by the introduction parts of the cassava harvester clamping device,this paper studies the feasibility of machine vision technology for the identification and positioning of cassava stalks in the field,which is of great significance for realizing the intelligent control of the cassava harvester holding device and the automation of the cassava harvester.The study in this paper is based on a single camera to identify and locate cassava stems by simulating the complex and varied environment in the field.First,field research and harvest experiments are carried out.The identification and positioning scheme of cassava stems and the required experimental equipment are determined.An indoor experimental platform is built.Enhance the acquired image and scale the cassava stem area using grayscale values method to highlight the cassava stem.The image is segmented using a fixed threshold method combined with morphological processing.The cassava stem area was selected using the pixel area method.A total of 9 parameters were extracted from the shape and texture characteristics of cassava stems and interferers,and the multi-layer perceptron classification model was trained.Using MATLAB to perform principal component analysis to determine the range of principal components,the grid search method combined with K-cross validation method was used to determine the number of hidden layer nodes of the multilayer perceptron was 9,and the number of principal components was 6.Using the intercepted 1602 cassava stalk pictures and 4759 pieces of interference images to carry out the transfer learning model training in the deep learning model RseNet50,it can solve the misjudgment problem of the multilayer perceptron.Indoors,under static conditions,the multi-layer perceptron recognition success rate is 92.5%,the algorithm average execution time is 0.51s,the average horizontal position error is 3.20mm,the average vertical position error is 2.53mm.The transfer learning model recognition success rate is 93.1%.The average execution time of the algorithm is 0.26 s,the average horizontal position error of the positioning is 3.03 mm,and the average vertical position error is 3.09 mm.Under dynamic and natural lighting conditions,the recognition success rate of both models is above 90%in the speed range of 0-0.1m/s.Dynamically and in dark environment,adjust the camera exposure time to 600μs with reinforcing light,at 0-0.2m/s speed,the success rate of both models is more than 90%.Using the Halcon Vision Library,QT Creator and Visual Studio,we designed a visual recognition software based on monocular camera.The software integrates camera calibration and recognition and positioning functions.It can be used and evaluated by many people to meet the needs of use.