Head Dermoscope Image Analysis Based On Target Detection Algorithm

Dermatoscopy,as the presentation of skin microstructure,plays an important role in the clinical diagnosis of diseases in dermatology.Dermatoscopy image analysis has become an important means of skin health assessment,and can also provide an important reference index for subsequent medical treatment.Clinical analysis of dermoscopy is a new technique developed in recent years.In the past,the application of dermoscopy in clinical diagnosis mainly focused on the morphological observation and clinical diagnosis of pigmentation lesions,while there were few reports on the observation and quantitative analysis of other morphological features of skin surface,especially hair features.Hair follicles,skin texture,dandruff,skin color and other information can be obtained through the observation of head dermoscope.In the dermoscope of the head,the number of hair follicles with normal growth is particularly important in the evaluation system of individual hair loss.If we rely on medical workers to manually identify and count hair follicles in the dermoscope of the head,it not only costs a lot of manpower,but also cannot guarantee the accuracy of manual identification and count of hair follicles due to the phenomenon of hair covering hair follicles in the dermoscope of the head,as well as redundant background information such as dandruff.Automatic recognition and counting of hair follicles is a target detection task in the field of image recognition.Target detection is divided into traditional target detection based on machine learning and target detection based on deep learning.Traditional target detection algorithms are characterized by manual construction,poor generalization and low automation.The target detection algorithm based on deep learning can automatically extract image features without human intervention,and has replaced the traditional target detection algorithm based on machine learning.Due to the phenomena of hair follicle occlusion and excessive redundant information in the data set of head dermoscopy,as well as the difference in the Angle and illumination of dermoscopy collection,the target detection algorithm based on deep learning will have some problems in the hair follicle detection task of head dermoscopy,such as missed detection,misdetection,and large detection frame positioning error.Through the analysis of several common target detection algorithms based on deep learning,this paper selects Faster R-CNN as the basis of hair follicle detection algorithm,and puts forward targeted solutions according to the problems existing in hair follicle detection by dermoscopy of the head.The main work of this paper is as follows:1.Analyses the current mainstream target detection algorithms based on the collected dermoscan images of the head,and selects the Faster R-CNN model with good comprehensive performance in combination with specific engineering requirements.2.through to the commonly used for target detection algorithm of feature extraction module convolution neural network analysis and experimental comparison,selection and quantity is less,the depth of the deeper residual network as the target detection algorithm of feature extraction,network characteristics of input to the model of original image are extracted,and add the space in the residual neural network attention mechanism and channel attention mechanism,to improve the quality of the network to extract the characteristics of the figure.3.For the RPN network part of the Faster R-CNN algorithm,K-Means clustering algorithm is introduced to cluster the bounding frames of hair follicles in the dermoscopy data set of the head,generating the anchor frame size suitable for the hair follicle inch size in the dermoscopy data set of the head,improving the quality of the generated candidate frames,and increasing the m AP value of the hair follicle detection algorithm from 0.7357 to 0.7405.Will Faster-R-CNN algorithm can improve the skin of the head mirror data sets,a experiment is carried out to analyze the experimental results shows:(1)due to residual network compared with ordinary convolution neural network has less parameters and deeper network level,the use of residual network for image feature extraction,at the same time to join in the residual network convolution attention mechanism module,was improved the detection accuracy of the model;(2)Clustering algorithm was introduced into the RPN network module of Faster R-CNN to generate the size and number of anchor frames suitable for hair follicle detection in the dermoscan data set of the head,which improved the detection accuracy of the model.