Plants play an important role in the entire ecosystem,with a global forest coverage rate of 31.7%and widespread.The identification and classification of plants can help us further improve the global ecosystem and establish a complete biological database.Plant leaves are the main reference object to distinguish their types.The point of entry of the traditional identification method lies in the morphological analysis of the contour and the trend of the vein.With the continuous innovation of modern technology,deep learning technology has gradually developed with its advantages in image feature extraction.This paper focuses on the model and implementation of plant leaf recognition algorithms based on improved deep learning.The main research contents include the following aspects:(1)The problem of overfitting when training a deep network model for small data samples is solved from two aspects.On the one hand,the method of data image preprocessing is researched and improved.The DCGAN method is used to expand the database image to meet the basic requirements for deep network training.On the other hand,transfer learning is used to apply the network model to parameter training to improve the accuracy of plant leaf recognition.The comparison experiments before and after optimization using Inception V3 and VGG-16 models show that this algorithm can improve the efficiency of identifying plant leaves.(2)Aiming at the problem that the network model parameters are too large,which leads to high storage and calculation costs,an improved neural network compression method is adopted.First,use recursive Bayes algorithm to perform network pruning to adaptively remove network redundancy;then,introduce K-means clustering to quantify convolutional layers and fully connected layers in the network.Though comparision experiments,the compression and accuracy of the three cases of only pruning,only quantization,and pruning plus quantization are compared.It is concluded that the fusion algorithm works best;Finally,combine classic convolutional neural network models VGG-16,for plant leaf recognition experiments.The results show that this algorithm can compress VGG-16 32 times.(3)The model was deployed on the embedded platform Jetson Nano,and the detection and recognition performance of the model was tested by the plant leaf images collected by the camera.Test results show that the model can recognize different kinds of leaves offline at a faster speed,and the recognition accuracy reaches 92.2%. |