Research On Image Recognition Of Ethnic Medicinal Plants Based On Deep Learning

With the rapid development and maturation of image recognition technology,the application of image recognition technologies,such as facial recognition,license plate recognition,and mobile QR code payment,has become increasingly widespread.However,in the field of medicine and pharmacology,image recognition technology for traditional Chinese medicine(TCM)plant identification,especially for ethnic medicine plant recognition,has not been widely adopted.Ethnic medicine is a part of my country’s traditional Chinese medicine industry,which has long been in a state of "keep in a boudoir without people knowing it".Using image recognition technology to identify ethnic medicines has important practical significance for the popularization of ethnic medicine-related knowledge.Based on deep learning-based image recognition methods,this paper proposes a recognition model for ethnic medicine plant images and develops an ethnic medicine plant image recognition system.Due to the special characteristics of ethnic medicine growth cycle and growth area,there are difficulties in collecting image data,which leads to a shortage of samples and complex background interference in the construction of ethnic medicine plant image datasets,making image recognition more challenging and affecting the recognition accuracy and classification performance of the model.To address these issues,this paper carries out a series of research works,including the following main contents:(1)Aiming at the problem of too small sample size in the task of ethnomedicine plant recognition,the image reconstruction algorithm model of ethnomedicine plant was constructed by combining multi-branch attention feature fusion network.The multi-span residual connection is used in the network backbone structure to reduce the time calculation burden in the training process.In the multi-branch feature fusion module,different channel information is interacted while the feature is refined.The cascaded attention block decouples the spatial channel attention from the channel attention so that the model can strengthen the image features from different angles.The optimized loss function can overcome the disturbance of outliers and accelerate the convergence of the model.Experiments have proved the effectiveness of this method.The reconstructed images of ethnic medicine plants have clear and natural senses,which can be used as data enhancement in image recognition.(2)Aiming at the problems of difficulty in feature extraction due to complex background of ethnomedicine plant images and overfitting of small samples in training,a two-channel adaptive attention neural network model combined with transfer learning strategy was proposed.The Tibetan RD dataset is constructed using the high-resolution ethnomedicine plant images reconstructed from the previous section to be added to the dataset.Different attention mechanisms were introduced into the extraction of high and low frequency features in the constructed model to make the network focus on the key features in the ethnomedicinal plant images with complex background.The original activation function is improved to make the model training more stable.Transfer learning is carried out on the model to alleviate the overfitting problem of small samples during training.The experimental results show that the proposed approach achieves better recognition accuracy and better training losses on the Tibetan RD dataset,and verifies the validity and generalization of the model over multiple publicly available image datasets.(3)Based on the above research,the image recognition system of ethnomedicine plants was developed by using Python language,which can identify ethnomedicine plant images and visualize the results to assist relevant personnel in the identification of ethnomedicine.