Research On Pollen Image Recognition Methods Based On Convolutional Neural Networks

Pollen identification plays an important role in fields such as palynology,forensic science,and paleoclimate reconstruction.At present,the recognition accuracy of most automatic pollen image recognition methods is low,and the training time of pollen image recognition models is long.Therefore,it is of great significance to establish a model that can identify pollen images with high accuracy and efficiency.Based on the public pollen image datasets POLEN23 E and POLLEN73 S,a convolutional neural network was introduced to identify pollen images,and the data enhancement,parameter optimization methods and the construction of pollen image recognition models were mainly studied.The main research contents are as follows:1)In order to ensure the reliability and authenticity of the experiment,the pollen image data are divided by random selection and K-fold cross verification,and the original image is processed by image flipping,color adjustment,Mix Up and Grid Mask,so as to suppress the over-fitting of the model,so as to improve the test accuracy of the model.2)To deal with the problem of low recognition accuracy caused by low resolution and similar shapes of pollen images,a pollen image recognition model based on dynamic and efficient network was proposed.First,the Noisy Student method is used to pre-train Efficient Net on the Image Net dataset;then,the trained weights are transferred to the pollen recognition model;finally,a dynamic learning rate is introduced to improve the recognition accuracy of the model.The simulation results show that the pollen image recognition model based on dynamic and efficient network has high classification accuracy.3)From the perspective of model training time,a pollen image recognition model based on improved residual network is proposed.First,fine-tune the Res Net50 model by adding Dropout layers and other methods;then,replace the standard convolution in the model with dilated convolution;Finally,replace the Relu activation function in the model with the Mish activation function.The simulation results show that the pollen image recognition model based on the improved residual network can use less training time and achieve satisfactory classification results.4)In order to better meet the needs of users for simplicity,speed and humanization.Therefore,this paper designs a pollen image recognition platform with better interactive effect.The platform software mainly includes five modules: user registration and login,pollen image data selection,pollen image recognition model training,pollen image recognition model testing and verification,and pollen image recognition.Through practical verification,the platform provides convenience for automatic identification of pollen images.Figure 35;Table 11;Reference 55...