Research And Application Of Chinese Herbal Medicine Leaf Recognition Algorithm Based On Deep Learning

Plants are one of the main forms of life on our planet and are an indispensable part of our lives.The classification and identification of plants allows us to better understand the characteristics of different types of plants,thus promoting the understanding and protection of plants.Some Chinese herbal medicines are also part of plants.The classification and identification of Chinese herbal medicine leaves can give us a more detailed understanding of the role of Chinese herbal medicines seen in daily life.The common Chinese herbal medicine plants can pass the texture of their leaf surface and external.The color of the leaves and the shape of the leaves themselves identify and classify the plants.Leaves are one of the most important organs of most plants.Common plants generally have leaves growing,and the leaves have characteristics that are easy to observe and preserve,and are often used as reference for plant identification.Moreover,most of the plant leaves tend to be flattened,relatively easy to collect,and the survival time is longer.The difference in texture and shape between the different types is obvious.Therefore,it is feasible to classify and identify plants based on the leaves.It is also a research hotspot of many researchers.According to the convolutional neural network weight sharing and local connection,and the input image has a certain degree of rotation,horizontal movement,invariance of the flip operation and so on.Based on the classical convolutional neural network model LeNet-5 model,this paper aims to improve the network structure and parameters of the original LeNet-5 model in the recognition of Chinese herbal plant image recognition,so that the improved model is identified.The Chinese herbal medicine leaf image has a higher recognition accuracy rate and a faster recognition rate.The specific work is as follows:(1)It summarizes the basic principles of CNN network,introduces the commonly used CNN model,LeNet-5 model,AlexNet model and the network structure of VGG model and their respective characteristics.Then introduces the commonly used activation functions of CNN network,and its advantages and disadvantages.The training process was described in detail.(2)In the process of classifying the input image for the traditional CNN network,the convolution block is used in the network structure to replace the traditional convolution-pooling structure.After convolution of the input image data in the convolutional neural network structure,the convolutional data is often further pooled.However,this operation often loses some input image information and destroys the convolution feature extraction.The layer extracts the feature information of the input image,and the feature information extracted by the single-layer convolution layer is not complete,and the method of superimposing the convolution layer can extract the feature information of the image more effectively and completely.The Chinese herbal medicine leaf recognition based on improved LeNet-5 network was proposed.The original convolution and pooling structure was improved by convolution block with three convolutional layers.The 3×3 small convolution kernel was used to improve the original model.The convolution kernel of ×5 and the activation function uses the ELU activation function instead of the sigmoid function,because the sigmoid activation function will cause the gradient to disappear when the output value is too large or too small;applying the global average pooling to prevent the fully connected layer in the network Over-fitting caused by too many parameters,and the Dropout operation in front of the fully connected layer can further prevent the over-fitting phenomenon of the network model;for the stochastic gradient descent algorithm SGD,it is initialized to the appropriate during the training process.The learning rate is relatively difficult and the SGD easily converges to the local optimal solution.The Adam algorithm is used instead of the original SGD algorithm.(3)In the linux system,the improved classification recognition model for Chinese herbal plant leaves is realized.The Python language and Django technology are used to implement the network model on the web side,establish the leaf image library,complete the classification identification and testing of the model,and adopt image normalization.The original Chinese herbal medicine leaf image was preprocessed by the method of thresholding,threshold segmentation,translation angle,rotation and flipping.The average recognition rate of the improved model reached 94.71%,which was higher than the traditional plant leaf recognition model.Compared with other plant leaf recognition models,the effectiveness of the proposed algorithm is verified.