CN114187495A - An Image-Based Apparel Trend Prediction Method - Google Patents

Abstract

The invention discloses a garment fashion trend prediction method based on images, which comprises garment image data acquisition, image foreground extraction, garment image feature extraction and garment fashion trend prediction. Firstly, collecting a clothing image data set, and preprocessing clothing image data; then obtaining a foreground image, and extracting clothing features based on a multi-convolution kernel deep neural network; and finally, a garment fashion trend prediction method based on deep learning is adopted, and the garment image characteristics are used as the input of the model to obtain the current garment fashion trend. The method can greatly reduce the calculation cost and the system complexity, promote the intellectualization of fashion trend prediction in the fashion field, and improve the effect and the quality of fashion prediction.

Description

An Image-Based Apparel Trend Prediction Method

technical field

The invention belongs to the technical field of intelligent clothing, and more particularly, relates to an image-based clothing fashion trend prediction method.

Background technique

At present, in the field of online clothing, designers usually design new clothing through their own knowledge and experience. It takes a lot of time and energy to design a piece of clothing each time, and it is impossible for designers to design every piece of clothing required. For a style of clothing, the clothing that will be popular in various regions in the future cannot be easily predicted, and usually requires the participation of multiple designers who are familiar with the region. Therefore, in the field of clothing, the intelligent prediction of the future development trend of clothing has potential and huge application scenarios.

The Chinese patent with publication number CN110705755A discloses "a method and device for predicting clothing trends based on deep learning", which collects pictures and information of popular clothing over the years from e-commerce clothing websites, performs feature extraction and integration, and then outputs according to the model results. The scheme of clothing rankings whose clothing popularity is topk, but this scheme is inaccurate for the prediction of clothing fashion trends and needs to be further optimized.

SUMMARY OF THE INVENTION

In view of the above defects or improvement needs of the prior art, the present invention provides an image-based clothing fashion trend prediction method, the purpose of which is to predict the clothing fashion trend through a deep learning method by collecting clothing pictures from major online clothing shopping websites. , reliable and real-time.

In order to achieve the above object, according to one aspect of the present invention, a method for predicting a fashion trend of clothing based on an image is provided, comprising the following steps:

Step 1, first collect the clothing image data set, and preprocess the clothing image data;

Step 2, using the foreground extraction model based on image multi-scale decomposition to extract the foreground of the clothing image;

Step 3, based on multi-convolution kernel deep convolutional neural network clothing image feature extraction and fusion, to obtain the final clothing image feature map;

Step 4, constructing a clothing fashion trend prediction model, and using the final clothing image feature as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend;

The clothing fashion trend prediction module includes: an adaptive weighted pooling layer, a fully connected layer and a softmax layer.

Further, in step 1, the clothing images of major shopping websites are collected by means of web crawlers and manual collection, wherein the shopping websites include Amazon-Online Shopping Mall, Tmall Mall, Taobao.com and Jingdong Mall.

Further, the preprocessing of the clothing image data includes: adjusting the image size through bilinear interpolation, and then performing image scale normalization and image normalization.

Further, the specific implementation of step 2 is as follows;

Step21, use total variation to decompose the image at multiple scales to obtain a series of smooth images;

Step22, express the foreground color distribution of the given smooth image as a Gaussian mixture model, and use the histogram shape analysis method to optimize the number of Gaussian functions of the Gaussian mixture model;

Step 23: Design iteration termination conditions according to the segmentation results of different smooth images, so that the foreground is extracted from the decomposition scale of the smooth images.

Further, the specific implementation of optimizing the Gaussian mixture model using the histogram shape analysis method in Step 22 is as follows;

The Gaussian distribution of the mth region is expressed as follows,

In the formula, G represents the Gaussian function, μ _m and ∑ m are the mean vector and covariance matrix of the color distribution in the region, respectively, u ( i ) represents the ith pixel in the smooth image u , and the pixel value is taken during calculation; det is Mathematical function for finding the determinant of a square matrix;

，直方图中共有256个值，分别用

表示，利用直方图波谷将图像分为N个区域，结合分割曲线计算出前景F 和背景B中区域个数，像素u(i)属于前景或背景的可能性为： Each peak of the histogram is used to represent the brightness distribution of the image area, and the median filter is used for smoothing. The smoothed histogram is

, there are 256 values in the histogram, respectively

means that the image is divided into N regions by using the histogram trough, and the number of regions in the foreground F and the background B is calculated by combining the segmentation curves. The possibility that the pixel u ( i ) belongs to the foreground or background is:

表示分割结果，其中，x _n =1表示前景，x _n =0表示 背景，L _F 和L _B 分别表示像素是u(i)前景和背景的可能性，ω _F 和ω _B 分别表示前景和背景的参 数，n _F 表示前景像素个数，n _B 表示背景像素个数；优化后高斯混合模型表示为： In the formula,

represents the segmentation result, where x _n =1 represents the foreground, x _n =0 represents the background, _LF and LB _represent the possibility that the pixel is u ( i ₎ foreground and background, respectively, ω F and ω _B represent the foreground and background, respectively The parameters of , n _F represents the number of foreground pixels, n _B represents the number of background pixels; the optimized Gaussian mixture model is expressed as:

Among them, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters.

Further, in Step 23, combining CrabCut and the color distribution model of the smooth image, the foreground extraction is converted into a joint optimization of segmentation and decomposition scale. The energy functional x* of foreground extraction is:

In the formula, α and β represent weights; the first term M ( u, u ⁰ ) is the multi-scale decomposition of the image, u ⁰ represents the original image, and u is the smooth image; the second term is the foreground extraction of the smooth image, S ( x, w,u ) is expressed as:

表示分割结果，其中，x _n =1表示前景，x _n =0表示背景；V(x, u)定义为将分割曲线放置在前景边界上的惩罚，如下式： In the formula, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters,

represents the segmentation result, where x _n =1 represents the foreground and x _n =0 represents the background; V ( x, u ) is defined as the penalty for placing the segmentation curve on the foreground boundary, as follows:

In the formula, A _i is the adjacent pixel set of the ith pixel, j is the pixel in A _i ; dis (∙) represents the Euclidean distance of the pixel pair; [∙] is the indicator function; γ and β represent the weight ; u(∙) denotes a smooth image.

Further, the multi-convolution kernel deep convolutional neural network includes a multi-convolution kernel feature extraction module and a multi-convolution kernel feature fusion module, and the specific implementation is as follows;

(31) Use the multi-convolution kernel feature extraction module to extract clothing image features, including style, color and style;

First, perform two convolution operations and activation function operations on the input image, extract the image features, and generate a clothing image feature map with a dimension of 224 × 224 × 64. The pooling operation converts the dimension of the feature map to 112×112×64; then, perform the convolution operation and the activation function operation, and perform the maximum pooling operation on the feature map extracted after the corresponding operation, and the generated dimension is 56×56 The feature map of ×128 is used as the input of the multi-convolution kernel fusion module;

(32) Use the multi-convolution kernel feature fusion module to fuse the extracted features to obtain the final clothing image feature map

The multi-convolution kernel feature fusion module includes: intra-module feature information fusion and inter-module feature information fusion;

Among them, the feature information fusion in the module includes three branches, and the three branches use convolution kernels with sizes of 3×3, 5×5 and 7×7 to further extract the features output by the multi-convolution kernel feature extraction module. The three branches are used to extract features in parallel, and the feature information fusion between modules is performed by aggregating the three branches in the feature information fusion within the module, and then performing a 3 × 3 convolution operation, and then inputting it to the feature information fusion part of the module again. Aggregate clothing feature information, and finally fuse the extracted feature information through a 1×1 convolution operation and aggregate it again with the input feature information to achieve the purpose of fusing the extracted feature information. In the above convolution operation, except Except for the 1×1 convolution operation, other convolution operations are followed by the Relu activation function.

Further, the specific processing process of the adaptive weighted pooling layer includes;

Input: feature map to be pooled, pooling window size n , loss function J , learning rate β ;

Step41: For each pooling layer, select the number of importance parameters according to the size of the pooling window of the layer, and randomly initialize n importance parameters k _i for the pooling window with n eigenvalues α _i , i =1 ,2,..., n ;

； Step42: Sort the eigenvalues in each pooling window in descending order to get

;

Step43: Perform softmax normalization on the initialized importance parameters to obtain weight parameters;

Step44: Multiply the weight parameter with the corresponding eigenvalue in each pooling window and accumulate to obtain the pooling result:

Step45: The initialized weight parameter w _i will be iteratively optimized through gradient descent with the progress of backpropagation during the training process until convergence:

Among them, α _i is the eigenvalue, ki is the random initialization parameter of the pooling window, wi _{is the} weight parameter, z is the weight parameter multiplied by the corresponding eigenvalue in each pooling window, and then accumulated to obtain the pooling result, ∂ represents partial differential.

Further, the loss function of the entire clothing fashion trend prediction model is the same as the loss function J in the adaptive weighted pooling layer, and the cross entropy loss function is used:

Among them, x represents the clothing image feature map of the input model, and p and q represent the true value of the clothing classification and the predicted value of the clothing classification, respectively.

Further, the processing process of the softmax layer is as follows;

Among them, Z _i is the output value of the ith node, and C is the number of output nodes, that is, the number of categories of the final classification result.

According to another aspect of the present invention, an image-based clothing fashion trend prediction system is provided, comprising the following modules:

The clothing image data collection module is used to collect clothing image data sets and preprocess the clothing image data;

The image foreground extraction module is used to extract the clothing image foreground using a foreground extraction model based on image multi-scale decomposition;

The clothing image feature extraction module is used for the clothing image feature extraction and fusion based on the multi-convolution kernel deep convolutional neural network to obtain the final clothing image feature map;

The clothing fashion trend prediction module is used to construct a clothing fashion trend prediction model, and the final clothing image feature is used as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend;

The clothing fashion trend prediction model includes: an adaptive weighted pooling layer, a fully connected layer and a softmax layer.

In general, compared with the prior art, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

(1) An image-based clothing fashion trend prediction method provided by the present invention uses clothing images to predict clothing fashion trends, including clothing styles, colors and styles, and predicts clothing fashion trends through a deep learning method, which is reliable and real-time;

(2) An image-based clothing fashion trend prediction method provided by the present invention, compared with the prior art, the present invention can greatly reduce the calculation cost and system complexity, and improve the effect and quality of fashion prediction.

Description of drawings

1 is a schematic flowchart of an image-based clothing fashion trend prediction system provided by an embodiment of the present invention;

FIG. 2 is a structural diagram of a deep convolutional neural network with multiple convolution kernels provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

As shown in Fig. 1, it is a kind of schematic flow chart of the clothing fashion trend prediction system based on image provided by the embodiment, including clothing image data collection module, image foreground extraction module, clothing image feature extraction module and clothing fashion trend prediction module 4 parts, The specific processing process of each module is as follows:

The clothing image data collection module is used to collect clothing image data sets and preprocess the clothing image data;

The image foreground extraction module is used to extract the clothing image foreground using a foreground extraction model based on image multi-scale decomposition;

The clothing image feature extraction module is used for the clothing image feature extraction and fusion based on the multi-convolution kernel deep convolutional neural network to obtain the final clothing image feature map;

The clothing fashion trend prediction module is used to construct a clothing fashion trend prediction model, and the final clothing image feature is used as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend;

The clothing fashion trend prediction model includes: an adaptive weighted pooling layer, a fully connected layer and a softmax layer.

Corresponding to the system, the present invention also provides an image-based clothing fashion trend prediction method, comprising the following steps:

(1) First collect the clothing image data set, and preprocess the clothing image data;

In this embodiment, the clothing images of major shopping websites are collected by means of web crawlers and manual collection, and image scale normalization and image standardization are performed on the clothing images, wherein the shopping websites include Amazon-online shopping mall (amazon.com), Tmall Mall (tmall.com), Taobao (taobao.com) and Jingdong Mall (jd.com).

Among them, the preprocessing of clothing image data includes: adjusting the image size by bilinear interpolation, and then performing image scale normalization and image normalization. In an embodiment, the image size is 224×224×3.

(2) Image foreground extraction, which is used to obtain foreground images;

The foreground of clothing images is extracted using a foreground extraction model based on image multi-scale decomposition.

Step21: Use total variation to decompose the image at multiple scales to obtain a series of smooth images. The decomposition protects the edge of the image, smoothes the texture, and compresses the color distribution range of the image area;

Step22: Express the foreground color distribution of the given smooth image as a Gaussian mixture model, and use the histogram shape analysis method to optimize the number of Gaussian functions of the Gaussian mixture model;

Among them, for each smooth image, the histogram shape analysis method is used to accurately model its color distribution.

Among them, the histogram shape analysis method optimizes the Gaussian mixture model. Assuming that the color distribution of each region in the smooth image is compact, the regional color distribution can be expressed as a Gaussian function, taking the mth region as an example:

In the formula, G represents the Gaussian function, μ _m and ∑ m are the mean vector and covariance matrix of the color distribution in the region, respectively, u ( i ) represents the ith pixel in the smooth image u , and the pixel value is taken during calculation; det is Mathematical function for finding the determinant of a square matrix;

，直方图中共有256个值，分别用

表示，利用直方图波谷将图像分为N个区域，结合分割曲线计算出前景F 和背景B中区域个数，像素u(i)属于前景或背景的可能性为： Each peak of the histogram is used to represent the brightness distribution of the image area, and the median filter is used for smoothing. The smoothed histogram is

, there are 256 values in the histogram, respectively

means that the image is divided into N regions by using the histogram trough, and the number of regions in the foreground F and the background B is calculated by combining the segmentation curves. The possibility that the pixel u ( i ) belongs to the foreground or background is:

表示分割结果，其中，x _n =1表示前景，x _n =0表示 背景，L _F 和L _B 分别表示像素是u(i)前景和背景的可能性，ω _F 和ω _B 分别表示前景和背景的参 数，n _F 表示前景像素个数，n _B 表示背景像素个数；优化后高斯混合模型表示为： In the formula,

represents the segmentation result, where x _n =1 represents the foreground, x _n =0 represents the background, _LF and LB _represent the possibility that the pixel is u ( i ₎ foreground and background, respectively, ω F and ω _B represent the foreground and background, respectively The parameters of , n _F represents the number of foreground pixels, n _B represents the number of background pixels; the optimized Gaussian mixture model is expressed as:

Among them, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters.

Step23: Design iteration termination conditions according to the segmentation results of different smooth images, so that the foreground is extracted from the decomposition scale of the smooth images.

被初始矩形框划 分为背景区域B和带有少量背景像素的前景区域F，RGB分别表示红绿蓝，u _R 表示红像素的集 合，u _G 表示绿像素的集合，u _B 表示蓝像素的集合，

表示原始图像u ⁰ 中所有像素的集合。其 前景提取的能量泛函x*为： Among them, the combination of CrabCut and the color distribution model of smooth images transforms foreground extraction into a joint optimization of segmentation and decomposition scales. an original image with N pixels

It is divided into a background area B and a foreground area F with a small number of background pixels by the initial rectangular frame, RGB represents red, green and blue respectively, u _R represents the set of red pixels, u _G represents the set of green pixels, u _B represents the set of blue pixels ,

represents the ^set of all pixels in the original image u0 . Its foreground extracted energy functional x* is:

In the formula, α and β represent weights; the first term M ( u, u ⁰ ) is the multi-scale decomposition of the image, u ⁰ represents the original image, and u is the smooth image; the second term is the foreground extraction of the smooth image, S ( x, w,u ) is expressed as:

表示分割结果，其中，x _n =1表示前景，x _n =0表示背景；V(x, u)定义为将分割曲线放置在前景边界上的惩罚，如下式： In the formula, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters,

represents the segmentation result, where x _n =1 represents the foreground and x _n =0 represents the background; V ( x, u ) is defined as the penalty for placing the segmentation curve on the foreground boundary, as follows:

In the formula, A _i is the adjacent pixel set of the ith pixel, j is the pixel in A _i ; dis (∙) represents the Euclidean distance of the pixel pair; [∙] is the indicator function; γ and β represent the weight ; u(∙) denotes a smooth image.

，其中〈∙〉表示均值。 In a specific embodiment, γ is taken as 50, in order to ensure that the energy of the above formula is larger when the gradient is low, and small when the gradient is high,

, where <∙> represents the mean.

(3) Extract and fuse clothing image features based on a multi-convolution kernel deep convolutional neural network to obtain the final clothing image feature map. The multi-convolution kernel deep convolutional neural network includes a multi-convolution kernel feature extraction module and a multi- Convolution kernel feature fusion module;

(31) Use a multi-convolution kernel feature extraction module to extract clothing image features, including style, color, and style.

In a specific embodiment, first, perform two convolution operations and activation function operations on the input image, extract the image features, and generate a clothing image feature map with a dimension of 224×224×64, and on this basis, extract the extracted image features. The feature image of the max pooling operation is performed to convert the feature map dimension to 112×112×64. Then, perform the convolution operation and activation function operation, and perform the maximum pooling operation on the feature map extracted after the corresponding operation to generate a feature map with a dimension of 56×56×128, which is used as the input of the multi-convolution kernel fusion module.

(32) Use the multi-convolution kernel feature fusion module to fuse the extracted features to obtain the final clothing image feature map.

As shown in FIG. 2 , it is a network structure diagram of a multi-convolution kernel feature fusion module provided by an embodiment, wherein the multi-convolution kernel feature fusion module includes: intra-module feature information fusion and inter-module feature information fusion.

Among them, the feature information fusion in the module includes three branches, and the three branches use convolution kernels with sizes of 3×3, 5×5 and 7×7 to further extract the features output by the multi-convolution kernel feature extraction module. The three branches are used to extract features in parallel, and the feature information fusion between modules is performed by aggregating the three branches in the feature information fusion within the module, and then performing a 3 × 3 convolution operation, and then inputting it to the feature information fusion part of the module again. Aggregate clothing feature information, and finally fuse the extracted feature information through a 1×1 convolution operation and aggregate it again with the input feature information to achieve the purpose of fusing the extracted feature information. In the above convolution operation, except Except for the 1×1 convolution operation, other convolution operations are followed by the Relu activation function.

(4) Construct a clothing fashion trend prediction model, and use the final clothing image features as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend.

Wherein, the clothing fashion trend prediction module includes: adaptive weighted pooling layer, fully connected layer and softmax.

Among them, the processing process of the adaptive weighted pooling layer specifically includes:

Input: feature map to be pooled, pooling window size n , loss function J , learning rate β ;

Step41: For each pooling layer, select the number of importance parameters according to the size of the pooling window of the layer, and randomly initialize n importance parameters k _i for the pooling window with n eigenvalues α _i , i =1 ,2,..., n ;

； Step42: Sort the eigenvalues in each pooling window in descending order to get

;

Step43: Perform softmax normalization on the initialized importance parameters to obtain weight parameters;

Step44: Multiply the weight parameter with the corresponding eigenvalue in each pooling window and accumulate to obtain the pooling result:

Step45: The initialized weight parameter w _i will be iteratively optimized through gradient descent with the progress of backpropagation during the training process until convergence:

Among them, α _i is the eigenvalue, ki is the random initialization parameter of the pooling window, wi _{is the} weight parameter, z is the weight parameter multiplied by the corresponding eigenvalue in each pooling window, and then accumulated to obtain the pooling result, ∂ represents partial differential.

Among them, the loss function of the entire clothing fashion trend prediction model is the same as the loss function J in the adaptive weighted pooling layer, and the cross entropy loss function is used:

Among them, x represents the clothing image feature map of the input model, and p and q represent the true value of the clothing classification and the predicted value of the clothing classification, respectively. The value obtained after the calculation of the cross entropy loss function does not necessarily meet the conditions and meaning of the probability distribution, so it is necessary to finally process the data into the form of a probability distribution through the softmax activation function to meet the multi-classification task requirements of the clothing image algorithm.

Among them, Z _i is the output value of the ith node, and C is the number of output nodes, that is, the number of categories of the final classification result.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (10)

1. an image-based clothing fashion trend forecasting method, is characterized in that, comprises the steps: Step 1, first collect the clothing image data set, and preprocess the clothing image data; Step 2, using the foreground extraction model based on image multi-scale decomposition to extract the foreground of the clothing image; Step 3, based on multi-convolution kernel deep convolutional neural network clothing image feature extraction and fusion, to obtain the final clothing image feature map; Step 4, constructing a clothing fashion trend prediction model, and using the final clothing image feature as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend; The clothing fashion trend prediction module includes: an adaptive weighted pooling layer, a fully connected layer and a softmax layer. 2. a kind of image-based clothing fashion trend prediction method as claimed in claim 1 is characterized in that: in step 1, collect the clothing images of major shopping sites by web crawler, manual collection mode, wherein shopping sites include Amazon- Online shopping mall, Tmall mall, Taobao and JD.com; The preprocessing of clothing image data includes: resizing the image by bilinear interpolation, and then performing image scale normalization and image normalization. 3. a kind of image-based clothing fashion trend prediction method as claimed in claim 1, is characterized in that: the concrete realization mode of step 2 is as follows; Step21, use total variation to decompose the image at multiple scales to obtain a series of smooth images; Step22, express the foreground color distribution of the given smooth image as a Gaussian mixture model, and use the histogram shape analysis method to optimize the number of Gaussian functions of the Gaussian mixture model; Step 23: Design iteration termination conditions according to the segmentation results of different smooth images, so that the foreground is extracted from the decomposition scale of the smooth images. 4. a kind of image-based clothing fashion trend forecasting method as claimed in claim 3, is characterized in that: in Step22, the concrete realization mode that utilizes histogram shape analysis method to optimize Gaussian mixture model is as follows; The Gaussian distribution of the mth region is expressed as follows,

In the formula, G represents the Gaussian function, μ _m and ∑ m are the mean vector and covariance matrix of the color distribution in the region, respectively, u ( i ) represents the ith pixel in the smooth image u , and the pixel value is taken during calculation; det is Mathematical function for finding the determinant of a square matrix; Each peak of the histogram is used to represent the brightness distribution of the image area, and the median filter is used for smoothing. The smoothed histogram is

, there are 256 values in the histogram, respectively

means that the image is divided into N regions by using the histogram trough, and the number of regions in the foreground F and the background B is calculated in combination with the segmentation curve. The possibility that the pixel u ( i ) belongs to the foreground or the background is:

In the formula,

represents the segmentation result, where x _n =1 represents the foreground, x _n =0 represents the background, _LF and LB _represent the possibility that the pixel is u ( i ₎ foreground and background, respectively, ω F and ω _B represent the foreground and background, respectively The parameters of , n _F represents the number of foreground pixels, n _B represents the number of background pixels; the optimized Gaussian mixture model is expressed as:

Among them, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters. 5. a kind of image-based clothing fashion trend forecasting method as claimed in claim 4 is characterized in that: in Step23, in conjunction with the color distribution model of CrabCut and smooth image, foreground extraction is converted into the joint optimization of segmentation and decomposition scale, foreground The extracted energy functional x* is:

In the formula, α and β represent weights; the first term M ( u, u ⁰ ) is the multi-scale decomposition of the image, u ⁰ represents the original image, and u is the smooth image; the second term is the foreground extraction of the smooth image, S ( x, w,u ) is expressed as:

In the formula, U ( x, w, u ) is the evaluation of the segmentation result of the smooth image u by x under the parameter w ; w represents the front and background color distribution parameters,

represents the segmentation result, where x _n =1 represents the foreground and x _n =0 represents the background; V ( x,u ) is defined as the penalty for placing the segmentation curve on the foreground boundary, as follows:

In the formula, A _i is the adjacent pixel set of the ith pixel, j is the pixel in A _i ; dis (∙) represents the Euclidean distance of the pixel pair; [∙] is the indicator function; γ and β represent the weight ; u(∙) denotes a smooth image. 6. a kind of image-based clothing fashion trend prediction method as claimed in claim 1 is characterized in that: described multi-convolution kernel depth convolutional neural network comprises multi-convolution kernel feature extraction module and multi-convolution kernel feature fusion module, the specific implementation is as follows; (31) Use the multi-convolution kernel feature extraction module to extract clothing image features, including style, color and style; First, perform two convolution operations and activation function operations on the input image, extract the image features, and generate a clothing image feature map with a dimension of 224 × 224 × 64. The pooling operation converts the dimension of the feature map to 112×112×64; then, perform the convolution operation and the activation function operation, and perform the maximum pooling operation on the feature map extracted after the corresponding operation, and the generated dimension is 56×56 The feature map of ×128 is used as the input of the multi-convolution kernel fusion module; (32) Use the multi-convolution kernel feature fusion module to fuse the extracted features to obtain the final clothing image feature map The multi-convolution kernel feature fusion module includes: intra-module feature information fusion and inter-module feature information fusion; Among them, the feature information fusion in the module includes three branches, and the three branches use convolution kernels with sizes of 3×3, 5×5 and 7×7 to further extract the features output by the multi-convolution kernel feature extraction module. The three branches are used to extract features in parallel, and the feature information fusion between modules is performed by aggregating the three branches in the feature information fusion within the module, and then performing a 3×3 convolution operation and then inputting it to the feature information fusion part of the module again. Aggregate clothing feature information, and finally fuse the extracted feature information through a 1×1 convolution operation and re-aggregate with the input feature information to achieve the purpose of fusing the extracted feature information. In the above convolution operation, in addition to Except for the 1×1 convolution operation, other convolution operations are followed by the Relu activation function. 7. a kind of image-based clothing fashion trend prediction method as claimed in claim 1 is characterized in that: the concrete processing process of adaptive weighted pooling layer comprises; Input: feature map to be pooled, pooling window size n , loss function J , learning rate β ; Step41: For each pooling layer, select the number of importance parameters according to the size of the pooling window of the layer, and randomly initialize n importance parameters k _i for the pooling window with n eigenvalues α _i , i =1 ,2,..., n ; Step42: Sort the eigenvalues in each pooling window in descending order to get

; Step43: Perform softmax normalization on the initialized importance parameters to obtain the weight parameters;

Step44: Multiply the weight parameter with the corresponding eigenvalue in each pooling window and accumulate to obtain the pooling result:

Step45: The initialized weight parameter w _i will be iteratively optimized through gradient descent with the progress of backpropagation during the training process until convergence:

Among them, α _i is the eigenvalue, ki is the random initialization parameter of the pooling window, wi _{is the} weight parameter, z is the weight parameter multiplied by the corresponding eigenvalue in each pooling window, and then accumulated to obtain the pooling result, ∂ represents partial differential. 8. a kind of image-based clothing fashion trend prediction method as claimed in claim 7 is characterized in that: the loss function of the whole clothing fashion trend prediction model is identical with the loss function J in the adaptive weighted pooling layer, and cross entropy is adopted. Loss function:

Among them, x represents the clothing image feature map of the input model, and p and q represent the true value of the clothing classification and the predicted value of the clothing classification, respectively. 9. a kind of image-based clothing fashion trend prediction method as claimed in claim 1, is characterized in that: the processing procedure of softmax layer is as follows;

Among them, Z _i is the output value of the ith node, and C is the number of output nodes, that is, the number of categories of the final classification result. 10. An image-based clothing fashion trend prediction system, characterized in that it comprises the following modules: The clothing image data collection module is used to collect clothing image data sets and preprocess the clothing image data; The image foreground extraction module is used to extract the clothing image foreground using a foreground extraction model based on image multi-scale decomposition; The clothing image feature extraction module is used for the clothing image feature extraction and fusion based on the multi-convolution kernel deep convolutional neural network to obtain the final clothing image feature map; The clothing fashion trend prediction module is used to construct a clothing fashion trend prediction model, and the final clothing image feature is used as the input of the clothing fashion trend prediction model to obtain the current clothing fashion trend; The clothing fashion trend prediction model includes: an adaptive weighted pooling layer, a fully connected layer and a softmax layer.

Images