Font Size: a A A

Research On The Recognition Method Of Herbal Tibetan Medicinal Materials Based On Deep Learning

Posted on:2024-09-13Degree:MasterType:Thesis
Country:ChinaCandidate:L Y ZhouFull Text:PDF
GTID:2544307085470734Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
As a key component of Tibetan medicine,correct identification and application of Tibetan medicinal herbs are important prerequisites for realizing their medicinal value.Due to the fact that most herbal Tibetan medicines grow in remote mountain areas that are difficult for ordinary people to access,their classification and identification are not well understood.The identification of Tibetan medicinal herbs mostly relies on manual methods,where professional technicians identify and authenticate the herbs by observing,touching,and smelling them.The accuracy of manual identification results heavily relies on the experience of the technicians,which can lead to errors.However,training professional technicians takes a long time,and the lack of talented individuals with rich experience in identifying Tibetan medicinal herbs is a major factor hindering the development of Tibetan medicinal herbs.Utilizing computer deep learning technology to achieve automatic identification of herbal Tibetan medicines under complex backgrounds(such as various stacked Tibetan medicinal herbs in natural scenes)is of great significance in promoting the modernization and internationalization of Tibetan medicinal herbs,as well as meeting the growing medical and healthcare needs of the Tibetan people and the development of the times.Based on this,this paper proposes a recognition model for herbal Tibetan medicine images using deep learning methods,focusing on three categories of herbal Tibetan medicines: sliced,elliptical-shaped,and plant.A series of research works were conducted around this topic,including:(1)There is currently no publicly available standard dataset for herbal Tibetan medicine image recognition,making it difficult to evaluate different methods under the same standards.To further expand the application scope of herbal Tibetan medicine image recognition,this paper constructed three standard complex background herbal Tibetan medicine datasets,including sliced herbal Tibetan medicine image dataset,ellipticalshaped herbal Tibetan medicine image dataset,and plant herbal Tibetan medicine image dataset.When establishing the dataset,factors such as the high similarity of herbal Tibetan medicine colors and shapes,strong light differences in shooting environments,high and low image quality,and different shapes and time spans of herbal medicine collection were fully considered.Data augmentation methods were used to address the problem of insufficient data from field shooting and web crawling.(2)To address the problem of low accuracy in existing automatic recognition of complex background sliced herbal medicines,this paper divided the RGB,HOG,and LBP features of herbal Tibetan medicine slices into several channels and used an improved HOG algorithm for multifeature fusion.The fused features were used as inputs to the Alex Net network model with an incorporated attention mechanism to recognize sliced herbal Tibetan medicine images in complex backgrounds.The proposed method was compared with existing Chinese patent medicine recognition methods,and the experimental results were analyzed.(3)Due to the high similarity of color and texture features in ellipticalshaped herbal Tibetan medicine images under complex backgrounds,texture features are key to distinguishing different types of ellipticalshaped herbal Tibetan medicines.This paper used Gabor wavelet transform and an improved LBP algorithm for texture feature extraction.The Gabor features encoded by the improved LBP algorithm were used as inputs to the Dense Net network,and Focal loss was used as the loss function.The effectiveness of the attention mechanism in this task was verified through experiments,and the proposed method was compared with existing methods for identifying similar-shaped medicinal herbs.(4)To address the complex growth environment of Tibetan medicinal plant in natural scenes and the fact that the same species of Tibetan medicinal plants may exhibit different morphological features at different growth stages,this paper studied the introduction of a multi-scale convolution and parallel dilated convolution feature extraction network with an attention mechanism,and combined it with navigator localization feature regions to construct a recognition model for Tibetan medicinal plant.The effectiveness of the proposed method was verified through experiments,and the proposed method was compared with existing methods for studying traditional Chinese medicinal plants.(1)Constructed three standard complex background herbal Tibetan medicine datasets,including 3,610 images of 32 sliced herbal Tibetan medicines,3,200 images of 18 elliptical-shaped herbal Tibetan medicines,and 8,963 images of 52 herbal Tibetan plant.All three datasets were expanded using data augmentation methods,and the effectiveness of data augmentation was verified through experiments.(2)Achieved recognition of sliced herbal Tibetan medicine images under complex backgrounds.The experimental results showed that multifeature fusion can effectively extract discriminative features of herbal Tibetan medicine slices under complex backgrounds.The proposed method of multi-feature fusion combined with deep learning achieved an image recognition accuracy of 91.68% for sliced herbal Tibetan medicines under complex backgrounds,which is better than other existing methods.This method also achieved good recognition results for sliced herbal Tibetan medicines with complex and heavily stacked backgrounds,showing promising application prospects.(3)Achieved recognition of elliptical-shaped herbal Tibetan medicine images under complex backgrounds.The experimental results showed that the improved LBP algorithm encoding Gabor features can more effectively extract the texture features of elliptical-shaped herbal Tibetan medicines.The proposed method achieved the highest recognition accuracy of 93.67% for elliptical-shaped herbal Tibetan medicine images in comparative experiments.The binary classification experiment of highly similar elliptical-shaped herbal Tibetan medicines proved that the proposed method can alleviate the problem of imbalanced dataset categories.(4)Achieved recognition of herbal Tibetan plant images under complex backgrounds.The experimental results showed that the feature extraction network constructed in this paper can effectively extract discriminative features of herbal Tibetan plants in natural scenes,and achieved high recognition rates for herbal Tibetan plants with complex and high similarity backgrounds.The proposed method achieved an image recognition accuracy of 94.1% for herbal Tibetan plant images,which is better than existing methods for identifying traditional Chinese medicinal plants.
Keywords/Search Tags:herbal, Tibetan medicine, deep learning, image recognition
PDF Full Text Request
Related items