Recognizing Of Microscopic Characteristics For Chinese Herbal Medicine Powder

Microscopic characteristics of Chinese herbal medicine powder have important effect and scientific value in identification.The microscopic images of Chinese herbal medicine powder have rich texture information and structural features,which can be used to effectively identify the Chinese herbal medicine powder.At present,most of the research aimed at small sample data and research under small-category conditions.It is poorly universal in large-scale Chinese herbal medicines,and difficult to form a universal identification method.Aiming at the problems about low accuracy and poor robustness of traditional classification and recognition algorithms,this paper proposes a method based on convolutional neural network for detecting and identifying Chinese herbal medicine powder by microscopic characteristics.Based on the microscopic image database of Chinese herbal medicine powder,the main innovations in this paper are summarized as follows:(1)A set of microscopic image database of Chinese herbal medicine powder is constructed to summarize the microscopic characteristics.Different types of microscopic images(gray,RGB and HSV)are researched.Three types of microscopic images are sent as training data to the convolutional neural network for experimental analysis,and various preprocessing methods are added to improve average precision.A variety of preprocessing methods(multi-light expansion,random cropping,multi-angle rotation,and scale standard normalization)are proposed to solve the problems of illumination effects,sample fragmentation,angular orientation,and scale size.Finally,the RGB images are chosen as the input of convolutional neural network,combining with a variety of image preprocessing methods.(2)Aiming at the microscopic catheter characteristic images,the region proposal network(RPN)of the Faster-RCNN algorithm is optimized.Combining with the attention mechanism of human eyes,the region proposal network(RPN)increases the weight of saliency feature in the microscopic images by adding pooling convolution layer and deconvolution layer.The scale parameter of anchors and loss function in the network are further optimized.The optimized Faster-RCNN algorithm increases the average precision of detection in microscopic catheter characteristic images from 74.50%to 82.14%.Its detection accuracy increases from 87.64%to 93.22%.(3)Microscopic characteristic images of Chinese herbal medicine powder are analyzed.The central of the region of interest ROI is intercepted as the fine-grained feature of microscopic characteristic images.Combining with the fine-grained feature of microscopic characteristic images,a two-channel improved SqueezeNet convolutional neural network is designed.The recognition results in two channels are fused by using probability weighting method.With optimized SqueezeNet convolutional neural network and the fine-grained feature of microscopic characteristic images,the results of experiment show that the correct identification rate of Chinese herbal medicine powder is improved to 90.33%.