Skin Disease Classification Using Mobilenet-RseSK Network

Abstract

The traditional deep learning method has a large amount of calculation, long time training and complex network structure, which is not easy to be applied to embedded or mobile devices. To solve the these problems, we proposed an improved lightweight mobile network named as Mobilenet-RseSK for skin disease classification. Firstly, a new attention mechanism seSK module is proposed, and seSK module is used to replace the position of SE in the original network. This module can better perform feature extraction and improve network performance than the original network attention module. Secondly, using RBN normalization, RBN maintains the advantages of BN, and strengthens the representation of specific features, which strengthens the degree of accuracy of skin disease identification. We compare the Mobilenet-RseSK network with MobilenetV3, Ghost and other advanced networks on the HAM10000 dataset. The proposed network promotes the accuracy of skin disease classification by 1.7% compared with the original network. Compared with advanced mobilenet network, our method achieves 85% accuracy on the test set. This network has certain practical value in skin disease classification, and is an effective lightweight skin disease classification algorithm.