| There are many types of skin diseases,but many diseases have similar appearances and colors,especially pigmented skin diseases.Usually,there are subtle differences between categories and large differences within categories.Ordinary people are less likely to identify different subcategories,and relying on domain experts is slow and has high labor costs,so the fine-grained image classification task is more difficult and deserving research than traditional large-scale image classification tasks.The convolutional neural network model performs well in large-scale image classification tasks,but its application in fine-grained classification needs to further research and train specific models for specific datasets.At the same time,mobile devices are gaining in popularity,mobile applications are developing rapidly,how to effectively integrate with deep learning technology is a new attempt.This thesis takes the dermoscopy images in the ISIC 2018 pigmented skin disease dataset as the research object,and develops a mobile terminal pigmented skin disease classification system based on lightweight convolutional neural network.First of all,analyze several classic convolutional neural network models such as VGG-16,Inception V3,Res Net-50,Xception,Mobile Net V1,Mobile Net V2 and so forth,using stratified k-fold cross validation model training and evaluation methods,evaluate models from the perspective of network model parameters,classification accuracy,sensitivity,and F1-score,and select lightweight convolutional neural network Mobile Net V2 with small parameters,accuracy,precision and F1-score values that are suitable as the basic network model of skin disease classification in this thesis.In the second place,an improvement project was proposed for the problem of insufficient dataset samples and uneven distribution of dataset samples in the skin disease dataset ISIC 2018.For the insufficient number of dataset samples,that can be resolved through data augmentation and transfer learning.A convolutional neural network training method based on online data augmentation and transfer learning algorithm is proposed.In view of the uneven distribution of samples,find solutions from the dataset and algorithm,compare the influences of different loss functions on the data model classification effects,design a new loss function based on Focal Loss to obtain a higher classification accuracy.Finally,a mobile application based on Android system is designed.Transplant the trained Mobile Net V2 network to the Android device,and complete the design of the mobile terminal system for skin disease classification,including a friendly human-computer interaction interface,database module,model calling module,and so on.The dermoscopy image input method can be selected from the album or real-time pictures,and the output result is the maximum probability value of the disease type. |
