多卷积神经网络模型融合的皮肤病识别方法

To solve the problem that the clinical features of basal cell carcinoma and seborrheic keratosis in skin diseases are very similar and difficult to classify, a multi-model fusion method of convolutional neural network (CNN) for dermatological recognition is proposed. The transfer learning method is used to train various CNN models, such as ResNet, Xception, and DensNet, to obtain the best recognition result for each model. Then, following the traditional fusion principle, voting and mean square error are considered as loss functions to fuse these three models to improve the recognition accuracy. To eliminate the influence of noise on skin disease recognition, and to heighten the accuracy and generalization ability of the proposed model, the maximum correntropy criterion (MCC) is used as the objective function of the multi-CNN fusion model, and the gradient ascent method is used to learn the contribution weight of different models according to the final results, thus a multi-CNN fusion model based on MCC is established. Experimental analysis is performed on the established basal cell carcinoma and seborrheic keratosis datasets. Compared with the prediction results of several single-model methods, the proposed multi-model fusion method achieves higher recognition accuracy. And compared with the traditional model fusion method, the proposed MCC-based multi-CNN fusion classification model gains strong generalization ability and can more effectively eliminate noise, it achieves an accuracy of 97.07%, exceeding that of the CNN single-model method and traditional multi-model fusion method.