The Human Feature Extraction And Measurement Based On Image

The improvement of science and technology has led to the development of the e-commerce industry.Online shopping has contributed to the creation of virtual fitting.However,errors in the extraction of feature points and the size measurement of human bodies have always been the problem,which make the effect of virtual fitting not ideal.In order to address this problem,we focus on the extraction of feature points and the size measurement of human body based on images,considering the situation of the single background and the complex background of target images.The key points of our research include:(1)We extract the features points and measure the size of human body by the improved method based on single background images.Firstly,it makes some preprocess for the single background images.Then,the human bounding box is detected based on the improved HOG feature,which narrows the search scope of the feature points.Finally,it solves the problem of partial absence of the profile of human body by some methods based on the proportion of human body to improve the measurement accuracy.(2)We segment human body images by simple-FCN model.Deep learning is applied to the image segmentation of complex background and the effect is significant.We propose the simple-FCN model by simplifying the structure of the existed FCN model,which composes of 7 convolution layers,5 pooling layers,3 deconvolution layers.Then the simple-FCN was applied to separate the human body from complex background in images.(3)We extract feature points and measure the size of human body by the improved ASM algorithm.Different from the traditional ASM algorithm with single matching template,we propose a method with multi matching templates.Firstly,the templates are classified into five classes according to the Euclidean distance between the face and body center,and the average template for each class is obtained from training.In the initial match,we use the Euclidean distance between two points to match the corresponding template,improving the accuracy of the match.Next,we set the feature point as center,select ten neighborhood points which located in the 3 rectangular regions,and make the target search in the gray model,so as to solve the problem that the traditional ASM method takes a long time to match and easily has the mismatch.Besides,based on the Mahalanobis distance,we select neighbor points which lie in 3 rectangular regions and match the gray values of these points with the gray model.Combining with the symmetry of the human body,we solve the problem that the fitting is only good for unilateral region of crotch.The experimental results show that a)the improved method locates the feature points of human body more accurately for single background images;b)The simple-FCN model segment the human body from the target image effectively;c)the improved ASM algorithm can adapt to the extraction of feature points and the size measurement of human body in images with complex background.Finally,this paper analyzes the shortcomings in the extraction of feature points and measurement of human body and puts forward further research directions.