2.1.

Experimental subjects

It can be found that most experimenters conduct the experiment through collecting a certain number of real body data on the basis of researching the previous two-dimensional non-contact body size prediction experiments. However, the problems of low sample coverage, high measurement cost and shortage of samples exist in terms of using the real body as the experiment object^{5, 6}, so that Make Human was considered to be used for manufacturing simulation body samples in the experiment. Make Human is the 3D character modeling software on the basis of numerous anthropological morphological characteristic data and face and limb models of male and female at different ages may be rapidly formed and accurately simulated.

The body model material was arranged as an Asian female in the experiment and the simulation body samples with body types of Y, A, B and C were randomly generated in accordance. Besides, simulation bodies of various body types were established through the careful modification of various parts of the simulation body, including slipped shoulder, square shoulder, hunchback, straight waist, abdominal retraction, abdominal bulge and so on. Finally, the data of various parts of the simulation body were exported. In the research, simulation bodies of 116 Asian females with the height between 150.3cm and 180.4cm, the weight between 40.72kg and 73.91kg and the age between 16 and 73 were the experiment objects. The basic size information of the experiment objects is shown in Table 1.

Table 1.

Basic size information of experiment objects.

2.2.

Measurement part determination

In the paper, the most widely available 5 body measurement sizes were adopted under default in terms of the body measurement data, namely the height, weight, chest circumference, waistline and hipline. The simulation body was arranged to stand naturally, look straight ahead, keep feet about shoulder-width apart and open hands diagonally downward and naturally at 30° to measure the height, weight, chest circumference, waistline and hipline of the simulation body.

2.3.

Two-dimensional image processing

The positive image was obtained for the simulation body sample and the Make Human software was called to open the simulation body model. The simulation body was shot positively through the virtual camera simulating the real scenes and the camera was arranged to be 1.5m above the ground with 2.5m-3.5m from the shot object. The whole simulation body was always kept in the viewfinder and the simulation body with bare feet stood in the center point of viewfinder. In addition, the posture of the simulation body should be consistent with the measurement posture and 5 positive images with the resolution ratio of 800×600 should be shot from different angles for the simulation body for the purpose of practicing the insensitivity to the shooting angles. Moreover, hands cannot cover the body in the shooting process in order to avoid the difficulty in obtaining data in the subsequent experiment process. Totally 580 positive images of the simulation body were obtained, which are shown in Figure 1a.

Figure 1.

Schematic diagram of image processing.

Binary image conversion was conducted for the positive images in order to more intuitively identify the body silhouette. Each pixel corresponds to a gray value in the positive image and the threshold function (threshold) was applied to the positive image to compare the gray value of the positive image with the threshold⁷. In addition, the simulation body was arranged to be black with the white background, thus achieving the binarization processing as shown in Figure 1b. Then the target silhouette was extracted through the edge detection. With local transition appearing in the gray values of two adjacent regions, the image edge appeared within two adjacent regions. Moreover, the boundary between the two regions may be obtained through the detection of the nearby edge points and the edge in the direction may be modeled through the location parameters on the directional profile. Besides, the first derivative and the second derivative were adopted to detect the edge in terms of the gray changes in the edge location⁸.

Baidu AI body key point recognition technology was used to generate the key point figure, which is shown in Figure 1c. 21 main key points of the simulation body in the positive image may be obtained by entering the positive image, containing the top of the head, five sense organs, the neck, limbs and other parts and the coordinate information of 5 key points may be extracted, such as the neck, the left shoulder, the right shoulder, the left hip and the right hip.

2.4.

Two-dimensional size extraction

The relevant two-dimensional image feature points (A: the foreneck point, B: the left lateral chest point, D: the right lateral chest point, E: the left lateral waist point, G: the right lateral waist point, H: the left lateral hip point and J: the right lateral hip point) should be determined in accordance with the obtained 5 body measurement sizes. The computer graphics knowledge was used to obtain the coordinate of various feature points, thus obtaining the pixel distance of various feature sizes, AC: neck-chest length, CI: chest-hip length, BD: lateral chest space, EG: lateral waist space and HJ: lateral hip space. In addition, the size extraction parts are shown in Figure 2.

Figure 2.

Schematic diagram of two-dimensional size extraction.

The coordinate information of A can be obtained directly through using the key point figure in accordance with Figure 2 and H and J can be found through the intersections of key points of the left and right hips in the key point figure respectively extending outwards to the silhouette edge. H and J were used to search along the silhouette edge from bottom to top until key points of left and right shoulders. The locations with the lowest horizontal distance from A were E and G. In addition, the locations with the maximum horizontal distance from A were B and D by searching along the silhouette edge from bottom to top and from E and G to key points of left and right shoulders. C, F and I were respectively the vertical intersections from A to BD, EG and HJ. Moreover, the feature sizes of 6 parts may be obtained by using the formula for distance between two points in accordance with the above principles. In addition, the real sizes of the body chest circumference, waistline and hipline in the image may be obtained through finding the conversion relationship between the image extraction size and the real size.

2.5.

Data pre-processing

The errors must exist in the process of obtaining the data and the pre-processing of the data should be conducted for the purpose of improving the accuracy of the data. SPSS software was used to process the missing value and find the abnormal value for correction of the original data. 580 effective samples were determined and the data were numbered. Then 485 groups of data were selected randomly to be used for the model training and the left 95 groups of data were used for the model verification.

3.1.

Establishment of the linear regression model

A linear regression model is a type of regression analysis that models the relationship between one or more independent and dependent variables using the least square function of a linear regression equation⁹. The linear regression model is featured by the advantages of simple calculation and the results easy to understand and the establishment steps of the linear regression model are as follows:

Linear Regression() in the machine learning function library was used to establish the linear regression model and the regression analysis was conducted for the training sample data. fit() was used to analyze the linear regression model parameters and predict () as the model established through the linear regression model parameters calculated by fit() was used to predict the input variables and obtain their values. Then the predict function was used to predict the chest circumference, waistline and hipline and the comprehensive analysis was conducted for their predicted values to determine the final predicted values. Then the model. intercept function was used to calculate the intercept and the model. coef function was used to calculate the coefficients of each parameter. The multiple linear regression equations of obtained chest circumference y₀, waistline y₁ and hipline y₂ are shown in Table 2.

Table 2.

Multiple linear regression equations.

2.2.

Establishment of MLP neural network model

As a type of artificial neural network model, multilayer perceptron model (MLP) can describe a group of complex mapping relationship¹⁰ from input variables to output variables through the nonlinear basis function. The establishment steps of MLP neural network model are as follows:

MLP Regressor() in the machine learning function library was called to conduct MLP neural network modeling for the chest circumference, waistline and hipline. Any one neuron at each layer of MLP will generate a connection coefficient with any one neuron at next layer and the output value of the output layer will be calculated in accordance with the connection coefficient after the input of the input layer, namely the predicted values of the chest circumference, waistline and hipline. In addition, they will be compared with the real values. The weight of various connection coefficients was adjusted with the times of repetition being the maximum number of the allowed iterations to obtain the neural network model meeting the requirements and then verify the model finally. In the paper, the arranged maximum iteration times were 10000 and the number of neurons of the input layer was 7, namely the number of the input vectors. Besides, the hidden layers were 2 with the number of neurons at the first hidden layer of 100 and the number of neurons at the second hidden layer of 50 and the number of the neurons of the output layer was 3, namely the number of the output variables. The basic structure of MLP neural network model is shown in Figure 3.

Figure 3.

Basic structure figure of MLP neural network model.

2.3.

Model evaluation

The index of R² value was adopted to evaluate the prediction accuracy and the degree of correlation between input vector and output vector of the above two models in the Paper. As a type of measurement, R2 value describes the difference between the predicted value and the actual value. In addition, when it is closer to 1, the prediction accuracy and the correlation are higher. The formula is as follows:

refers to the actual value and refers to the predicted value with indicating the mean in the formula. The specific numerical values were obtained through calculation and the results of the verified model are shown in tables 3 and 4.

Table 3.

Evaluation results of linear regression model.

Table 4.

Evaluation results of MLP neural network model.

In accordance with Tables 3 and 4, the correlation coefficients of the chest circumference, waistline and hipline in the linear regression model and MLP neural network model all exceed 0.8 and the average error, standard variance, maximum error and minimum error of the linear regression model are all less than those of the MLP neural network model. As a consequence, the prediction accuracy of the linear regression model is better than that of the MLP neural network model and the prediction accuracy difference of the two models is minor.