Research And Implementation Of Skin Disease Classification Algorithm Based On Depth Feature Fusion

Skin cancer is one of the most common cancers at present.In terms of detection and treatment,it is of positive significance to study the algorithm to distinguish the specific types of lesions in skin disease images.In view of the rapid development of deep learning,convolutional neural network has made a great breakthrough in skin disease image processing.However,the skin disease recognition algorithm still has room for improvement due to the problems of large similarity between skin diseases,small difference within skin diseases and limited image samples.This paper mainly solves the above problems by enhancing the feature extraction ability of the network and fusing multi-modal features.(1)In order to enhance the feature extraction ability of convolutional neural network and improve the multi classification recognition rate,D-F-Efficient Net network is proposed in this paper.The main components of the network are Efficient Net-B5,improved SPP network,multi-layer feature fusion module and ECA attention network.The improved SPP network uses pooled cores of different sizes,which can increase the receptive field and obtain rich semantic information;The multi-layer feature fusion module reduces the loss of underlying lesion information through feature reuse;ECA attention network learns the channel weight relationship to improve the difference between features.The experimental results on ISIC 2019 data set show that the BACC value of D-F-Efficient Net network is 4.5%higher and the ACC value is 3.4% higher than that of Efficient Net network.The experimental results show the effectiveness of the D-F-Efficient Net network proposed in this paper.(2)In order to solve the problem that the number of dermatological image samples is small and the available information is insufficient,this paper participates in the network training in the form of text,and proposes a multi-modal feature fusion network.The network includes an improved gated attention unit and feature stitching network,which can combine the complementary information of different modal data to provide multi-level lesion features to improve the recognition rate.The experimental results on PAD-UFES-20 data set show that the BACC value of this network is 3.3% higher than that of the existing Meta Block network,It is 9.8%higher than the BACC value of concatenation network.The experimental results show that the multi-modal network proposed in this paper is effective.Finally,according to the software engineering specification,this paper completes the design and development of skin disease auxiliary diagnosis system.Implement the research results and provide users with online diagnosis services of skin diseases.