Research And Implementation Of Skin Lesions Recognition Based On Micro Attention Mechanism

The skin is one of the important organs in the human body and also plays a protective role in other organs.With people’s living standards and the construction of modern cities,people’s living environment has become increasingly polluted,resulting in a gradual increase in the incidence rate of skin diseases,and the number of skin cancer patients is also increasing year by year.At present,the diagnosis of skin lesions in the medical field mainly depends on the dermatology’s manual observation or observation of the shape,color and other characteristics of the lesions with the help of dermatoscope images,and makes the corresponding diagnosis based on experience.Due to the high similarity between some skin lesions,even experienced dermatology may misjudge,and there is some uncertainty.With the development of artificial intelligence and deep learning technology,computer-aided diagnosis of skin lesions in smart healthcare will become a future trend.At present,there are also some difficulties in automatic recognition of skin lesion images.If the location of skin lesions is different,the boundary between the lesion area and normal skin is not clear enough,and there is high similarity between different lesions or other benign skin features,automatic recognition will face certain difficulties.In response to the above issues,this article analyzes the models and methods for skin lesion image recognition.In order to improve recognition quality and focus on small differences in features,the focus is on the recognition method of attention mechanism.At the same time,to meet the requirements of applying CNN models to intelligent environments with limited computing power such as mobile devices,a lightweight convolutional neural network is fused with attention mechanism to propose a skin lesion recognition model based on micro attention mechanism.The main tasks completed in this article are as follows:(1)This paper systematically studies the characteristics of traditional convolutional neural networks and lightweight convolutional neural networks,combines the complex characteristics of human skin feature extraction,integrates Efficient Net convolutional neural networks with attention mechanisms,and conducts comprehensive optimization in multiple dimensions such as model layer depth,layer width,and input resolution,focusing on improving the efficiency and accuracy of the model.In the main network module MBConv,compression and excitation optimization are added,and the input activation graph is expanded using 1 × 1 convolution to increase the depth of the feature graph.In order to consider the attention characteristics between channels,an improved Efficient Net ECA-Efficient Net V2 model is proposed,which achieves efficient and accurate recognition tasks.(2)In order to enhance the feature extraction ability of convolutional neural networks,GAMCA＿Efficient Net V2 network model was proposed based on the accuracy and efficiency of fine grained skin classification.The network is mainly composed of Efficient Net V2,CA attention mechanism,and GAM global attention mechanism modules.The Efficient Net V2 model is a smaller and faster new series that can adaptively adjust regularization and image size compared to the prototype Efficient Net through training for NAS awareness and scaling discovery.The CA attention mechanism learns the positional weight relationships of channels to improve the differences between features.For the task of fine grained image classification,the global attention mechanism GAM is used to extract higher order information from images.The experimental results on the dataset show that the GAMCA＿Efficient Net V2 network has significantly improved both the Train＿acc and the Test＿acc compared to the original Efficient Net network.The experiments show that it improves the accuracy and also accelerates the training speed.The results show that the GAMCA＿Efficient Net V2 network proposed in this article is accurate and efficient.(3)In order to solve the problem of a small number of skin disease image samples and insufficient information available,this article will integrate the ISIC2019 and Derm Net data sets on the basis of the ISIC2018 skin lesion data set,which to some extent solves the problem of uneven data types and expands the range of skin disease types that can be diagnosed and referenced.Finally,based on the GAMCA＿Efficient Net V2 network,the design and development of a skin lesion recognition system was completed,providing support for online auxiliary diagnosis and intelligent medical treatment of skin lesions...