Research On The Technology Of Ginseng Appearance Quality Grading Based On Deep Learning

Chinese medicine is a unique medicine,moreover,the core of Chinese medicine,whose unique drug composition and pharmacological properties provide the basis for the therapeutic tools of Chinese medicine.The unique status and role that Chinese medicine possesses in Chinese medicine is also an important marker that separates Chinese medicine from other medical sciences.Throughout millennia of struggle,Chinese medicine has made remarkable achievements in the treatment of various diseases and has had a significant impact both at home and abroad.At the same time,Chinese medicine is also of great significance in the history and cultural development of mankind as an important cultural heritage.However,traditional algorithms are inadequate in identifying the appearance and quality of ginseng,and the rate of manual identification is slow and inefficient.Therefore,this paper combines the advantages of deep learning in image recognition classification,selects raw sun ginseng as the research object,proposes a deep learning-based ginseng appearance and quality grading method,and combines the improved model with pruning operation to compress the This paper proposes a deep learning-based ginseng appearance quality classification method,and combines the improved model with pruning operation to compress the model,so as to reduce the number of model ginseng and provide a basis for the subsequent mobile application of ginseng appearance quality classification.The main research work in this thesis is as follows:(1)Construction and processing of the dataset: four types of ginseng samples of different grades were photographed using smartphones and professional photography studios,and this dataset was augmented with data to expand the dataset and finally construct a standard dataset of ginseng images.(2)Construction of a ginseng appearance quality grading recognition model: the traditional activation function Re LU is replaced by the Leaky Re LU activation function to enhance the expression capability of the model;the ECA channel attention mechanism module is introduced on the residual block to improve the sensitivity of the model to the ginseng pixels and to obtain the feature information of ginseng more accurately;the focus loss function is introduced to balance the dataset and combined with the The model is trained with the idea of migration learning.(3)Model compression and mobile deployment: The model is compressed using pruning methods to optimize the parameters of the model.The mobile deployment uses the Android studio integrated development tool to achieve the recognition of ginseng appearance and quality in a fast and efficient way.In this paper,the improved model was subjected to multiple ablation experiments on the ginseng dataset,and the results showed that the improved Res Net50 mainly outperformed other models in terms of accuracy and loss values,and it also showed greater advantages in terms of convergence speed and stability,with slightly faster training time per round than other models.Compared with the classical convolutional neural network models VGG16,Google Net,Res Net50 and Densenet121,the improved Res Net50 has the best performance,with an accuracy of 97.39% and a loss value of 0.035 in the test set.The training time per round was also slightly faster than the other models.The trained model was then pruned and successfully deployed to mobile.In summary,the method can improve the recognition accuracy without compromising the appearance of ginseng,and can basically meet the recognition requirements.