Research On Microscopic Feature Identification For Chinese Medicinal Materials Powder Based On Multi-channel

Microscopic image identification of Chinese medicinal materials has scientific and practical value as an important means to guarantee the quality of Chinese medicinal materials.Most of the commonly used microscopic image recognition methods have the problems of few identification types and poor robustness.Deep learning,as a mainstream image processing method,has excellent self-learning ability and can meet the complex computational requirements.Therefore,it is of research significance to combine deep learning with microscopic feature images of Chinese herbal medicine powders to realize a microscopic feature identification method that is easy to extend,low-cost and professional.The main work and innovation of the paper are as follows:(1)Complete the image generalization of microscopic features of powdered Chinese herbs under the intersection of pharmacology and image science,and realize the classification of 14 microscopic features as well as the establishment of 100 kinds of Chinese herbs data set.In order to eliminate the impact of data randomness problem on network learning in image acquisition,image pre-processing methods such as random rotation,shearing and normalization are used to realize the expansion of microscopic feature position and texture randomness as well as the unification of illumination,so as to improve the network’s learnable ability on data.(2)For the problem of multi-scale features and fuzzy features of microscopic features,the design and optimization of coarse detection structure is realized based on YOLO v3 network.The multi-scale training is based on the full convolutional network of YOLO v3,and the network is trained randomly on the images through five scales to achieve the learning of the differences between features at different scales and the improvement of the detection accuracy of multi-scale features under large scale images.Secondly,a local context(LC)module is proposed in the backbone network of YOLO v3 to improve the detection effect of blurred features by adding different scale averaging pooling layers to achieve the focus on local edge and texture information under the global information integration.Experiments show that multi-scale training achieves a 2.8% improvement and adding LC module on top of multi-scale training achieves a 2.0% improvement,bringing the m AP to 71.2%.(3)An improved refined classification method combining Res Net v2 network and attention mechanism is proposed for the problem of classifying many kinds of Chinese herbal medicines and small samples.At the input side of the network,multi-color space fusion combines the original RGB color space and HSV color space in parallel to form a six-channel network input method,so that the HSV color space,which is more compatible with human eye vision,can realize the network to supplement the information of small samples in terms of hue,saturation and luminance.Inside the network,the attention mechanism uses the global average pooling(GAP)layer and the global maximum pooling layer in the channel attention module to achieve the compression and fusion of information in the channel dimension through the fully connected layer;and the GAP layer in the spatial attention module to achieve the extraction and fusion of information in the spatial dimension through the convolution of three different size voids.Thus,the weights of channel and spatial dimensions are assigned more effectively in the training process to achieve accurate classification under multiple types of small samples.Experiments show that the multi-color spatial fusion approach achieves a 1.8% improvement in recognition rate,the attention mechanism approach achieves a 3.1% improvement in recognition rate,and the combination of both achieves a 4.1% improvement,resulting in a 93.9% recognition rate.Through the structure of coarse detection and fine classification,the recognition,data query and database functions of Chinese herbal medicines are realized,and the construction of the microscopic image recognition system of Chinese herbal medicines powder features is completed.