Research On Microscopic Image Identification For Chinese Medicinal Materials Powder Based On Fine-Grained

The identification of traditional Chinese medicine samples by the method of microscopic image recognition of traditional Chinese medicine powder is a key technology for its quality control,and it is of great significance to carry out this research work.In this paper,deep learning is used to establish a Fine-Grained microscopic image identification framework for Chinese medicinal materials,that is,convolutional neural networks are used to fully learn subtle discriminative features,and finally a general computer Chinese medicinal material powder microscopic image recognition method is realized.The main work and innovation of the paper are as follows:(1)A detection method based on improved information fusion and attention mechanism is proposed.Firstly,the PANet part of the YOLO v4 target detection algorithm is improved to make good use of the small target features.Secondly,the CBAM module is added between the main feature extraction network of YOLO v4 and PANet to make the model better focus on foreground features.The experimental results show that after improving PANet,the m AP can be improved by 2.4%.On this basis,the CBAM module can be added to achieve a 0.6% increase.Finally,the m AP can reach 83.2%.For small-scale targets such as starch granules and crystallization,obtain a good detection result.(2)A classification method based on improved dynamic Re LU activation function and attention mechanism is proposed.First,an improved SE-Net is added to the residual structure of the Xception network,so that it can better focus on the object itself in the image and suppress useless information.Secondly,the improved dynamic Re LU activation function is used in the Xception network to allow different samples has their own adaptive activation function slope in order to better fits the training data.The experimental results show that in terms of recognition rate,only the improved SE-Net is improved by 1.1%,only the improved dynamic Re LU activation function is improved by 1.2%,and the recognition rate of the combination of the two improvements is comprehensively improved by 3.1%.(3)A classification method based on multi-channel fusion and improved SPP-Net structure is proposed.First,the Canny edge detection feature map and the local binary pattern feature map are combined with the original RGB image at the input to form a five-channel input image,which is used to increase the dimension of the input data on the channel to provide richer texture information.Second,the network structure In this paper,SPP-Net is improved and embedded into the Efficient Net to increase the depth of the network,so that it can mine the deep texture information of the image without being affected by the acquisition environment and solve the problem of image recognition across databases.The experimental results show that the recognition rate of this method has increased by 2.7%,reaching 81.5%.