Research On Fungal Image Classification Algorithm Based On Deep Learning

The level of medical care in China has increased year by year.However,the detection methods of fungi are relatively weak,mainly relying on manual morphological identification,which is difficult to meet clinical needs.With the continuous development of computer computing speed and artificial intelligence technology,deep learning is the most popular machine learning method at present,which can perform large-scale data representation learning.The paper studies the automatic classification of filamentous fungi images based on deep learning techniques.For the classification problem of filamentous fungi image classification,the paper adopts two methods of target detection and image classification.Firstly,a YOLO target detection model based on deep learning is established to detect the target conidial position and its classification.The experimental results show that the YOLO model has poor detection effect on small targets or adjacent targets,and is easily affected by complex background background,resulting in target frame positioning.Inaccurate question.In addition,based on the artificially extracted HOG features,the SVM model is establlished to classify fungal images,and the classification accuracy is higher than that of the YOLO model.However,the data preparation time is long and it is difficult to achieve real-time classification.Aiming at the shortcomings of YOLO algorithm and SVM classification,this paper establishes a deep learning-based CNN network model to extract deep features of fungal images.Considering that the original fungal image contains multiple conidia,and the background is complex,there are bubbles and other interference factors.The fungal image preprocessing method effectively extracts all target conidia in the image;in order to meet the requirements of deep learning for a large number of training samples,a data enhancement method with arbitrary angle rotation is proposed to expand the sample by nearly 30 times.In addition,the Dropout design of the CNN network effectively avoids network over-fitting,and the experimental verification data enhancement and Dropout technology can effectively improve the model classification effect.Finally,for the over-fitting problem of image classification,a fungal image classification model based on migration learning is established.The CNN is preprocessed on the ImageNet dataset to obtain the network model parameters,and the feature extraction ability of the network is ensured.Train again,fine-tune other parameters of the network,and finally classify the fungal images using the softmax classifier.The experimental results show that the migration learning strategy can effectively avoid over-fitting,improve the accuracy of the algorithm,and make the network have good robustness.In order to facilitate clinical data management,this paper carried out the design of fungal automatic classification management information system,mainly designed database E-R diagram,data dictionary and visualization interface,which provided reference for hospital fungal detection,improving doctor's work quality and work efficiency.