Skin Disease Recognition Method Based On ResNeXt

With the continuous expansion of the incidence of skin diseases,the scope of the disease continues to increase,skin disease diagnosis technology has attracted more and more attention from researchers.Traditional technology relies on dermatologist’s artificial diagnosis method,which is easily affected by subjective factors,and it is difficult to ensure the accuracy of diagnosis.With the rapid development of computer technology,computer-aided diagnosis and treatment have gradually become a powerful helper for doctors.In recent years,the computeraided diagnosis and treatment method represented by deep learning has achieved good performance in the field of skin disease image recognition,but there are still some shortcomings,such as: when the convolution neural network of a single model starts from the whole image,the focus degree of the lesion area is not enough,it is easy to be affected by the background image,and the key features of skin disease are improved The low efficiency and the difficulty of global optimization of the fusion strategy with fixed weight value in the multimodel fusion method affect the effect of skin disease recognition in varying degrees.In order to solve the above problems,based on the ResNeXt model,this paper improves the single model recognition and multi-model fusion recognition and proposes the following two skin disease recognition methods:(1)In single model recognition,a method of dermatoscopy image recognition based on Attention-ResNeXt is proposed.On the basis of ResNeXt,an attention mechanism is introduced,and three Attention-ResNeXt sub-models are constructed: CA-ResNeXt based on channel attention mechanism,SA-ResNeXt based on spatial attention mechanism,and MAResNeXt based on mixed attention mechanism.The experiment compares the recognition effect of three Attention-ResNeXt sub-models and non-attention ResNeXt model in the same skin disease recognition task from the accuracy,accuracy,recall,and F1 score.The experimental results show that the Attention-ResNeXt model with attention mechanism performs better in skin disease image recognition task than the ResNeXt model without attention mechanism in the sub-model,CA-ResNeXt has the best recognition effect,and the recognition accuracy reaches 86.27%,which is better than the other two attention mechanisms.(2)In multi-model fusion recognition,a weight adaptive multi-convolution fusion method is proposed.This method is compatible with the current mainstream convolutional neural network model and has good diversity and flexibility.In the process of training,according to the real-time performance of each model,the weight of the model can be dynamically adjusted by using the weight adaptive algorithm,so as to achieve the best recognition state.The experimental results show that the recognition effect of the multi-model fusion method is generally better than that of the single model recognition method;the proposed weight adaptive multi convolution fusion method has better recognition performance than other multi-model fusion methods,and in the weight adaptive multi convolution fusion method,the model combination of AlexNet + VGG16 + ResNet18 + ResNeXt50 has the best recognition effect,and the recognition accuracy reaches 98%.