Research On Pigment Separation And Segmentation Algorithms For Skin Images

The human skin,located on the surface of the human body,is an important factor affecting the appearance of the body and an important barrier against disease.Separation of skin pigments refers to the extraction of relative content and distribution data of pigments,including two major pigments in skin: melanin and hemoglobin,through the analysis of color images of human skin.Dermoscopy image segmentation is designed to detect the location and edge contour of skin lesions.With the continuous development of computer technology,combining with computer analysis of skin pigment distribution,it has a good auxiliary role in cosmetics and medical industry.At the same time,computer aided diagnosis is widely used in the diagnosis and automatic evaluation of dermoscopy and clinical images,which is of great significance to the treatment of skin diseases.Based on skin optical model,independent component analysis was used to obtain melanin and hemoglobin information.In this thesis,the algorithm of skin pigment separation is improved from the perspective of data screening.In recent years,with the continuous maturity of deep learning,dermoscopy image segmentation gradually uses deep learning network model to solve the problem of image segmentation and obtain the contour of skin lesions.In this thesis,the semantic segmentation network process was optimized and the segmentation accuracy was improved by combining the skin pigment separation method.On the basis of the existing research status,this thesis mainly carries out the following work:(1)A sub-block filtering algorithm based on image channel distance was proposed.In order to make the separation model of skin pigment more accurate and clearer,a subblock filtering algorithm was proposed to filter the input single skin image data.The input image is divided into sub-blocks of the same size,and the sub-blocks are sorted and screened according to the average pixel difference between channels,and the input data of the pigment separation model is recombined.(2)A local clustering dimensionality reduction method for image data was proposed.To solve the problem of unstable calculation process in the method of independent component analysis(ICA),local clustering and dimensionality reduction were carried out on the preprocessed data to reduce the data scale and computational complexity.After data preprocessing,a certain number of clustering centers are uniformly set,and the clustering centers are redistributed according to the Euclidean distance between the sample points and the clustering center.At the end of clustering,a certain number of sample points are randomly selected according to the clustering center to continue the calculation of pigment separation.Experimental results show that the stability of the algorithm is significantly improved.(3)A Segnet network segmentation method combining skin pigment separation was proposed.Based on the framework of deep learning algorithm,the skin mirror image was segmtioned by combining the pigment separation algorithm and the Segnet network model of semantic segmentation network.Firstly,pigments are separated from the dermoscopy images to obtain the corresponding melanin and hemoglobin images.Then,the melanin and hemoglobin images are transformed into single-channel grayscale images,which are combined with the original images and expanded into image data with channel number of 5.Then,the extended image is segmented by Segnet deep neural network in semantic segmentation network.Experimental results on ISIC-2018 dermoscopy image data set show that the proposed algorithm achieves a good segmentation effect under the premise of comprehensive consideration of accuracy,sensitivity and specificity.