Automatic diagnosis of melanoma image based on improved YOLOv5
An improved YOLOv5 model was proposed as a countermeasure to the problems of insufficient recognition accuracy,uneven samples and insufficient lightweight of hair occlusion targets in existing melanoma intelligent diagnosis models.Firstly,the CS_Neck structure was designed based on the improved C3 structure and self-attention mechanism,so as to effectively distinguish the related features of melanoma and hair.Secondly,a method of mining difficult samples with secondary screening was proposed,in which focus loss function was used to reduce the weight of simple samples,and the idea of loss rank mining(LRM)was intro-duced to reduce the number of simple samples.Finally,the lightweight backbone network was designed,and the improved RepVGG structure was proposed to replace the common convolutional extraction features,improve the inference speed,and the width multiplier was introduced to reduce the number of parameters and weights so as to realize the lightweight model.The experi-mental results,based on ISIC2019 data set,show that the weights and parameters of the proposed algorithm are only 7.9 MB and 4.0×106,and the accuracy reaches 92.9%.The proposed algorithm can effectively improve the accuracy and achieve lightweight,indicating good meeting of the requirements of efficient diagnosis of melanoma.
melanoma detectionYOLOv5attention mechanismmining hard sampleslightweight
NETL
NSTL
万方数据
