An intelligent product selection system that can be adjusted based on real-time analysis of data
Through the real-time analysis and intelligent product selection system based on data, the problem of online shopping and product selection is solved, and the efficient and accurate product selection process and improved product quality are achieved.
Patent Information
- Application Number
- CN202410900370.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-05
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2044-07-05
AI Technical Summary
In the prior art, online shopping anchors spend time and difficult to guarantee quality during product selection, and there is a situation where the selected products are of poor quality.
The real-time analysis and intelligent product selection system based on data is adopted to obtain platform product information through the data acquisition module, the data processing module predicts product demand, and the quality judgment module determines the hardness of the product to be selected, and determines the target product.
It improves product selection efficiency and quality, accurately predicts product demand, and improves the efficiency and accuracy of product quality judgment.
Smart Images
Figure CN118863583B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, specifically to the field of intelligent product selection systems, and in particular to an adjustable intelligent product selection system based on real-time analysis of data. Background Art
[0002] With the advancement of electronic technology and the development of the Internet, online shopping has become a mainstream shopping method. Currently, each anchor needs to hire a large number of people to select products. The product selection process is time-consuming and time-consuming. In addition, due to people's subjective feelings during the selection process, the quality of the selected products cannot be guaranteed, and there is a situation where the quality of the selected products is actually poor. Summary of the invention
[0003] The embodiments of the present invention provide an intelligent inventory management system and method, electronic device and storage medium based on a big data supply chain, which can perform real-time analysis based on data and intelligently select products, thereby improving the efficiency and quality of product selection.
[0004] On the one hand, the present invention provides an intelligent product selection system and system that can be adjusted based on real-time analysis of data, the system comprising: a data acquisition module: used to obtain platform product information in the most recent period, the platform product information including at least one of the following: product click volume, product purchase volume, product purchase volume and product search volume; a data processing module: used to predict the product demand for the next period based on the platform product information; determine a preset number of product types based on the product demand; a data acquisition module, which obtains N products to be selected corresponding to each product type, where N is a positive integer; a quality judgment module, which is used to judge the quality of the N products to be selected, and determine the target product based on the quality judgment result; The method of performing quality judgment on the N commodities to be selected and determining the target commodity based on the quality judgment result includes: the shape of the N commodities to be selected is a bar; the quality judgment process of the N commodities to be selected is filmed by a camera to obtain a commodity shooting video, and the quality judgment process includes: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; swinging the target commodity to be selected based on a preset swing amplitude and swing angle; determining a first video frame and a second video frame at a first angle and a second angle from the commodity shooting video, respectively, the first angle and the second angle being two angles corresponding to the maximum swing amplitude; identifying the target commodity to be selected in the first video frame and the second video frame; determining M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; the curvatures at T first feature points and T second feature points are calculated based on the first curve and the second curve, respectively, and the T first feature points and T second feature points are the feature points farthest from the clamping device among the M first feature points and the M second feature points; a first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; the hardness of the target commodity to be selected is determined based on the first average curvature and the second average curvature; the target commodity is determined based on the hardness of each of the N commodities to be selected.
[0005] In one possible implementation, the embodiment of the present application can obtain the demand / sales volume of a certain product category in the next cycle based on a comprehensive prediction of the product click volume, product purchase volume, product add-on volume and product search volume of the product category, and select a preset number of product types with high demand / sales volume as the final product category.
[0006] For example, the product category may specifically be grilled sausages, and the cycle is one week. In this application, the system may determine the demand for each product category based on the product click volume, product purchase volume, product add-on volume, and product search volume of each product category in the past week; if the grilled sausage category ranks high, determine that the grilled sausage category is the type of product that needs to be selected; the N products to be selected may be the grilled sausages corresponding to N brands respectively; the target product may be one or more of the N products to be selected.
[0007] Preferably, the commodity to be selected can be food; the hardness of some foods reflects the quality and taste of the food, and foods that are too hard or too soft are not of the best quality and taste. Therefore, by calculating the hardness, foods with appropriate hardness can be selected; the quality and taste of the video can be judged more accurately, thereby improving the quality of product selection.
[0008] Preferably, the hardness of some products indicates the durability of the products. For example, a ruler that is too hard will break easily, and a ruler that is too soft will deform easily. A ruler with a higher hardness is neither easy to break nor deform, reflecting the best product quality.
[0009] It can be seen that the embodiment of the present application acquires and analyzes data through the data acquisition module and the data processing module, which can accurately predict the demand for commodity categories and improve the accuracy of commodity category selection; the embodiment of the present application judges the quality of commodities through the quality judgment module, which improves the efficiency of commodity quality judgment during commodity selection and improves the accuracy of commodity quality judgment.
[0010] Preferably, determining the target commodity based on the hardness of each of the N commodities to be selected comprises: inputting an image of each of the N commodities to be selected into a trained target neural network, and outputting a quality grade of each of the N commodities to be selected; obtaining a score of each of the N commodities to be selected based on the hardness and the quality grade of each of the N commodities to be selected; and determining the target commodity based on the score of each of the N commodities to be selected.
[0011] Preferably, the target neural network includes: an encoder, a first feature fusion module, a second feature fusion module, a third feature fusion module, and a decoder; the encoder includes five feature extraction modules, which extract the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map in order from large to small scales; and the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map are input into the corresponding convolution layer compression channel to obtain the corresponding first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map; wherein the first feature map contains underlying texture features, and the fifth feature map contains high-level semantic features; the first feature map, the second feature map, and the fifth feature map are input into the first feature fusion module to obtain a first fused feature map; the first feature map The feature map, the third feature map and the fifth feature map are input into the second feature fusion module to obtain the second fused feature map; the first feature map, the fourth feature map and the fifth feature map are input into the third feature fusion module to obtain the third fused feature map; the decoder includes the first saliency module, the second saliency module and the third saliency module; the third fused feature map is input into the third saliency module to obtain the first intermediate feature map; the second fused feature map is fused with the first intermediate feature map and then input into the second saliency module to obtain the second intermediate feature map; the first fused feature map is fused with the second intermediate feature map and then input into the first saliency module to obtain the saliency map; supervised training is performed based on the first intermediate feature map, the second intermediate feature map, the saliency map and the true value map based on the target loss function to obtain a trained target neural network.
[0012] It should be noted that as the number of convolution layers increases, texture features gradually disappear and semantic features gradually increase, that is, the bottom-level features contain more texture features, and the high-level features contain more semantic features. The complexity and variability in the product recognition process, the use of texture features and semantic features at the same time can achieve better recognition results; the target neural network can detect the saliency map of the target in the image. The saliency map integrates texture information and semantic information. The saliency map can highlight the texture information of the target, thereby improving the accuracy of quality judgment.
[0013] Preferably, the first feature map, the second feature map and the fifth feature map are input into the first feature fusion module to obtain the first fused feature map, including: the first feature fusion module includes a first positioning module and a first texture module; the first positioning module reduces the dimension of the fifth feature map and the second feature map by average pooling respectively, and then inputs two fully connected layers with activation functions to generate a first channel attention map, and then uses the first channel attention map and the second feature map to perform channel attention calculation to obtain a first positioning map; in the first texture module, the first positioning map is upsampled and a first intermediate positioning map and a second intermediate positioning map are generated by a self-attention mechanism; the first feature map is downsampled and a first attention feature map is generated by a self-attention mechanism, and the upsampled first positioning map has the same dimension as the downsampled first feature map; a first correlation between the position in the first intermediate positioning map and the texture feature in the first attention feature map is determined, and based on the first correlation, the texture feature in the first attention map is fused into the second intermediate positioning map to obtain a first texture map to present the texture of the feature; the first texture map is fused with the upsampled first positioning map by a residual connection to obtain a first fused feature map.
[0014] Preferably, the first feature map, the third feature map and the fifth feature map are input into the second feature fusion module to obtain the second fused feature map, including: the second feature fusion module includes a second positioning module and a second texture module; the second positioning module reduces the dimension of the fifth feature map and the third feature map by average pooling respectively, and then inputs two fully connected layers with activation functions to generate a second channel attention map, and then uses the second channel attention map and the third feature map to perform channel attention calculation to obtain a second positioning map; in the second texture module, the second positioning map is upsampled and a third intermediate positioning map and a fourth intermediate positioning map are generated by a self-attention mechanism; the first feature map is downsampled and a second attention feature map is generated by a self-attention mechanism, and the upsampled second positioning map has the same dimension as the downsampled first feature map; the second correlation between the position in the third intermediate positioning map and the texture feature in the second attention feature map is determined, and based on the second correlation, the texture feature in the second attention map is fused into the fourth intermediate positioning map to obtain a second texture map to present the texture of the feature; the second texture map is fused with the upsampled second positioning map by residual connection to obtain a second fused feature map.
[0015] Preferably, the first feature map, the fourth feature map and the fifth feature map are input into the third feature fusion module to obtain the third fused feature map, including: the third feature fusion module includes a third positioning module and a third texture module; the third positioning module reduces the dimension of the fifth feature map and the fourth feature map respectively by using average pooling, and then inputs two fully connected layers with activation functions to generate a third channel attention map, and then uses the third channel attention map and the fourth feature map to perform channel attention calculation to obtain a third positioning map; in the third texture module, the third positioning map is upsampled and the fifth intermediate positioning map and the sixth intermediate positioning map are generated by the self-attention mechanism; the first feature map is downsampled and the third attention feature map is generated by the self-attention mechanism, and the upsampled third positioning map has the same dimension as the downsampled first feature map; the third correlation between the position in the fifth intermediate positioning map and the texture feature in the third attention feature map is determined, and the texture feature in the third attention map is fused into the sixth intermediate positioning map based on the third correlation to obtain a third texture map to present the texture of the feature; the third texture map is fused with the upsampled third positioning map by using residual connection to obtain a third fused feature map.
[0016] Preferably, the loss function is:
[0017]
[0018] in and They are BCE loss function and IoU loss respectively. GT is a binary truth map with a value range of (0,1). is the output of the i-th saliency module, i=1,2,3 Indicates upsampling.
[0019] On the other hand, an embodiment of the present invention provides an intelligent product selection method that can be adjusted based on real-time analysis of data, the method comprising: obtaining platform product information in the most recent period, the platform product information comprising at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume; predicting product demand in the next period based on the platform product information; determining a preset number of product types based on the product demand; N products to be selected corresponding to each product type, where N is a positive integer; performing quality judgment on the N products to be selected, and determining a target product based on the quality judgment result; The method of performing quality judgment on the N commodities to be selected and determining the target commodity based on the quality judgment result includes: the shape of the N commodities to be selected is a bar; the quality judgment process of the N commodities to be selected is filmed by a camera to obtain a commodity shooting video, and the quality judgment process includes: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; swinging the target commodity to be selected based on a preset swing amplitude and swing angle; determining a first video frame and a second video frame at a first angle and a second angle from the commodity shooting video, respectively, the first angle and the second angle being two angles corresponding to the maximum swing amplitude; identifying the target commodity to be selected in the first video frame and the second video frame; determining M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; the curvatures at T first feature points and T second feature points are calculated based on the first curve and the second curve, respectively, and the T first feature points and T second feature points are the feature points farthest from the clamping device among the M first feature points and the M second feature points; a first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; the hardness of the target commodity to be selected is determined based on the first average curvature and the second average curvature; the target commodity is determined based on the hardness of each of the N commodities to be selected.
[0020] In another aspect, an embodiment of the present invention provides an electronic device, comprising: a processor, adapted to implement one or more instructions; and a computer storage medium, wherein the computer storage medium stores one or more instructions, wherein the one or more instructions are adapted to be loaded by the processor and execute the following steps:
[0021] Obtain platform product information in the most recent period, the platform product information including at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume; predict product demand for the next period based on the platform product information; determine a preset number of product types based on the product demand; N products to be selected corresponding to each product type, where N is a positive integer; perform quality determination on the N products to be selected, and determine a target product based on the quality determination result; The method of performing quality judgment on the N commodities to be selected and determining the target commodity based on the quality judgment result includes: the shape of the N commodities to be selected is a bar; the quality judgment process of the N commodities to be selected is filmed by a camera to obtain a commodity shooting video, and the quality judgment process includes: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; swinging the target commodity to be selected based on a preset swing amplitude and swing angle; determining a first video frame and a second video frame at a first angle and a second angle from the commodity shooting video, respectively, the first angle and the second angle being two angles corresponding to the maximum swing amplitude; identifying the target commodity to be selected in the first video frame and the second video frame; determining M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; the curvatures at T first feature points and T second feature points are calculated based on the first curve and the second curve, respectively, and the T first feature points and T second feature points are the feature points farthest from the clamping device among the M first feature points and the M second feature points; a first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; the hardness of the target commodity to be selected is determined based on the first average curvature and the second average curvature; the target commodity is determined based on the hardness of each of the N commodities to be selected.
[0022] In another aspect, an embodiment of the present invention provides a computer storage medium, characterized in that the computer storage medium stores one or more instructions, and the one or more instructions are suitable for being loaded by a processor and executing the following steps:
[0023] Obtain platform product information in the most recent period, the platform product information including at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume; predict product demand for the next period based on the platform product information; determine a preset number of product types based on the product demand; N products to be selected corresponding to each product type, where N is a positive integer; perform quality determination on the N products to be selected, and determine a target product based on the quality determination result; The method of performing quality judgment on the N commodities to be selected and determining the target commodity based on the quality judgment result includes: the shape of the N commodities to be selected is a bar; the quality judgment process of the N commodities to be selected is filmed by a camera to obtain a commodity shooting video, and the quality judgment process includes: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; swinging the target commodity to be selected based on a preset swing amplitude and swing angle; determining a first video frame and a second video frame at a first angle and a second angle from the commodity shooting video, respectively, the first angle and the second angle being two angles corresponding to the maximum swing amplitude; identifying the target commodity to be selected in the first video frame and the second video frame; determining M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; the curvatures at T first feature points and T second feature points are calculated based on the first curve and the second curve, respectively, and the T first feature points and T second feature points are the feature points farthest from the clamping device among the M first feature points and the M second feature points; a first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; the hardness of the target commodity to be selected is determined based on the first average curvature and the second average curvature; the target commodity is determined based on the hardness of each of the N commodities to be selected.
[0024] It can be seen that the present application acquires and analyzes data through the data acquisition module and the data processing module, and can accurately predict the demand for commodity categories, thereby improving the accuracy of commodity category selection; the embodiment of the present application judges the quality of the commodity through the quality judgment module, thereby improving the efficiency of commodity quality judgment during commodity selection, and improving the accuracy of commodity quality judgment. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.
[0026] Figure 1 It is a scenario architecture diagram provided by an embodiment of the present invention.
[0027] Figure 2 It is a flow chart of an adjustable intelligent product selection system based on real-time analysis of data provided by an embodiment of the present invention.
[0028] Figure 3 Schematic diagram of the target neural network architecture.
[0029] Figure 4 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0030] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0031] It should be noted that the terms "first", "second", etc. in the description and claims of the embodiments of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order.
[0032] like Figure 1 As shown, Figure 1 This is a scenario architecture diagram provided in the embodiment of the present application. The embodiment of the present application provides an intelligent product selection system that can be adjusted based on real-time analysis of data and can be applied in the following situations: Figure 1In the scene architecture shown, there are a clamping device A1 and a target commodity A2; specifically, there are three states in which the clamping device A1 clamps the target commodity A2, namely, a stationary state 101, a state 102 swinging to one side to an angle of β, and a state 103 swinging to the other side to an angle of β; the swing amplitude of the clamping device is 2β degrees; the camera can capture the swinging process of the clamping device A1 clamping the target commodity A2, when in the stationary state 101, 5 feature points are selected on the target commodity A2, and the distances between the 5 feature points are the same; when in the state 102, 5 feature points on the target commodity A2 in the image are determined, and based on the The curve is obtained by fitting the positions of the five feature points; the curvature of the two feature points farthest from the clamping device A1 is calculated based on the curve, wherein the two farthest feature points are the two feature points at the top in the static state 101; the first average curvature is calculated based on the curvature of the two farthest feature points; when in state 103, the five feature points on the target product A2 to be selected in the image are determined, and the curve is obtained by fitting the positions of the five feature points; the curvature of the two feature points farthest from the clamping device A1 is calculated based on the curve; the second average curvature is calculated based on the curvature of the two farthest feature points; the hardness of the target product is calculated based on the first average curvature and the second average curvature. It can be understood that the greater the curvature, the lower the hardness, and the smaller the curvature, the greater the hardness. .
[0033] Based on the above description, the embodiment of the present invention proposes an intelligent product selection system that can be adjusted based on real-time analysis of data. The intelligent product selection system that can be adjusted based on real-time analysis of data can be executed by an electronic device. Figure 2 The intelligent product selection system that can be adjusted based on real-time analysis of data has the following modules:
[0034] S201, data acquisition module: used to obtain platform product information in the most recent period, wherein the platform product information includes at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume.
[0035] S202, data processing module: used to predict the demand for goods in the next cycle based on the platform goods information; determine a preset number of goods types based on the demand for goods;
[0036] S203, a data acquisition module is further used to acquire N to-be-selected commodities corresponding to each commodity type, where N is a positive integer;
[0037] S204, a quality determination module, configured to determine the quality of the N selected commodities and determine the target commodity based on the quality determination result;
[0038] Specifically, the quality judgment of the N commodities to be selected and the determination of the target commodity based on the quality judgment result include: the shape of the N commodities to be selected is a bar; the quality judgment process of the N commodities to be selected is filmed by a camera to obtain a commodity shooting video, and the quality judgment process includes: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, M≥2; the M feature points are equal to each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; swinging the target commodity to be selected based on a preset swing amplitude and swing angle; determining a first video frame and a second video frame at a first angle and a second angle from the commodity shooting video, respectively, the first angle and the second angle being two angles corresponding to the maximum swing amplitude; identifying the target commodity to be selected in the first video frame and the second video frame; determining M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; the curvatures at T first feature points and T second feature points are calculated based on the first curve and the second curve, respectively, and the T first feature points and T second feature points are the feature points farthest from the clamping device among the M first feature points and the M second feature points; a first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; the hardness of the target commodity to be selected is determined based on the first average curvature and the second average curvature; the target commodity is determined based on the hardness of each of the N commodities to be selected.
[0039] In one possible implementation, the embodiment of the present application can obtain the demand / sales volume of a certain product category in the next cycle based on a comprehensive prediction of the product click volume, product purchase volume, product add-on volume and product search volume of the product category, and select a preset number of product types with high demand / sales volume as the final product category.
[0040] For example, the product category may specifically be grilled sausages, and the cycle is one week. In this application, the system may determine the demand for each product category based on the product click volume, product purchase volume, product add-on volume, and product search volume of each product category in the past week; if the grilled sausage category ranks high, determine that the grilled sausage category is the type of product that needs to be selected; the N products to be selected may be the grilled sausages corresponding to N brands respectively; the target product may be one or more of the N products to be selected.
[0041] Optionally, the product to be selected may be food; the hardness of some foods reflects the quality and taste of the food, and foods that are too hard or too soft are not of the best quality and taste. Therefore, by calculating the hardness, foods with appropriate hardness can be selected; the quality and taste of the video can be judged more accurately, thereby improving the quality of product selection.
[0042] Optionally, the hardness of some products indicates the durability of the product. For example, a ruler that is too hard will break easily, and a ruler that is too soft will deform easily. A ruler with a higher hardness is not easy to break or deform, which reflects the best product quality.
[0043] It can be seen that the embodiment of the present application acquires and analyzes data through the data acquisition module and the data processing module, which can accurately predict the demand for commodity categories and improve the accuracy of commodity category selection; the embodiment of the present application judges the quality of commodities through the quality judgment module, which improves the efficiency of commodity quality judgment during commodity selection and improves the accuracy of commodity quality judgment.
[0044] In one possible implementation, the determining the target product based on the hardness of each of the N products to be selected includes: inputting the image of each of the N products to be selected into a trained target neural network, and outputting the quality grade of each of the N products to be selected; obtaining the score of each product to be selected based on the hardness and the quality grade of each of the N products to be selected; and determining the target product based on the score of each of the N products to be selected.
[0045] In one possible implementation, see Figure 3 , Figure 3Schematic diagram of the architecture of the target neural network; the target neural network includes: an encoder, a first feature fusion module, a second feature fusion module, a third feature fusion module, and a decoder; the encoder includes five feature extraction modules, which extract the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map in order from large to small scales; and the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map are input into the corresponding convolution layer compression channel to obtain the corresponding first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map; wherein the first feature map contains underlying texture features, and the fifth feature map contains high-level semantic features; the first feature map, the second feature map, and the fifth feature map are input into the first feature fusion module to obtain the first fused feature map ; Input the first feature map, the third feature map and the fifth feature map into the second feature fusion module to obtain the second fused feature map; input the first feature map, the fourth feature map and the fifth feature map into the third feature fusion module to obtain the third fused feature map; the decoder includes a first saliency module, a second saliency module and a third saliency module; input the third fused feature map into the third saliency module to obtain a first intermediate feature map; fuse the second fused feature map with the first intermediate feature map and input them into the second saliency module to obtain a second intermediate feature map; fuse the first fused feature map with the second intermediate feature map and input them into the first saliency module to obtain a saliency map; perform supervised training based on the first intermediate feature map, the second intermediate feature map, the saliency map and the true value map based on the target loss function to obtain a trained target neural network.
[0046] It should be noted that as the number of convolution layers increases, texture features gradually disappear and semantic features gradually increase, that is, the bottom-level features contain more texture features, and the high-level features contain more semantic features. The complexity and variability in the product recognition process, the use of texture features and semantic features at the same time can achieve better recognition results; the target neural network can detect the saliency map of the target in the image. The saliency map integrates texture information and semantic information. The saliency map can highlight the texture information of the target, thereby improving the accuracy of quality judgment.
[0047] In a possible implementation, the first feature map, the second feature map and the fifth feature map are input into a first feature fusion module to obtain a first fused feature map, including: the first feature fusion module includes a first positioning module and a first texture module; the first positioning module uses average pooling to reduce the dimension of the fifth feature map and the second feature map respectively, and then inputs two fully connected layers with activation functions to generate a first channel attention map, and then uses the first channel attention map and the second feature map to perform channel attention calculation to obtain a first positioning map; in the first texture module, the first positioning map is upsampled and a first intermediate positioning map and a second intermediate positioning map are generated using a self-attention mechanism; the first feature map is downsampled and a first attention feature map is generated using a self-attention mechanism, and the upsampled first positioning map has the same dimension as the downsampled first feature map; a first correlation between a position in the first intermediate positioning map and a texture feature in the first attention feature map is determined, and based on the first correlation, the texture feature in the first attention map is fused into the second intermediate positioning map to obtain a first texture map to present the texture of the feature; the first texture map is fused with the upsampled first positioning map using a residual connection to obtain a first fused feature map.
[0048] In a possible implementation, the first feature map, the third feature map and the fifth feature map are input into the second feature fusion module to obtain the second fused feature map, including: the second feature fusion module includes a second positioning module and a second texture module; the second positioning module uses average pooling to reduce the dimension of the fifth feature map and the third feature map respectively, and then inputs two fully connected layers with activation functions to generate a second channel attention map, and then uses the second channel attention map and the third feature map to perform channel attention calculation to obtain a second positioning map; in the second texture module, the second positioning map is upsampled and a third intermediate positioning map and a fourth intermediate positioning map are generated using a self-attention mechanism; the first feature map is downsampled and a second attention feature map is generated using a self-attention mechanism, and the upsampled second positioning map has the same dimension as the downsampled first feature map; the second correlation between the position in the third intermediate positioning map and the texture feature in the second attention feature map is determined, and based on the second correlation, the texture feature in the second attention map is fused into the fourth intermediate positioning map to obtain a second texture map to present the texture of the feature; the second texture map is fused with the upsampled second positioning map using a residual connection to obtain a second fused feature map.
[0049] In a possible implementation, the first feature map, the fourth feature map and the fifth feature map are input into a third feature fusion module to obtain a third fused feature map, including: the third feature fusion module includes a third positioning module and a third texture module; the third positioning module reduces the dimension of the fifth feature map and the fourth feature map by using average pooling, and then inputs two fully connected layers with activation functions to generate a third channel attention map, and then uses the third channel attention map and the fourth feature map to perform channel attention calculation to obtain a third positioning map; in the third texture module, the third positioning map is upsampled and a fifth intermediate positioning map and a sixth intermediate positioning map are generated by a self-attention mechanism; the first feature map is downsampled and a third attention feature map is generated by a self-attention mechanism, and the upsampled third positioning map has the same dimension as the downsampled first feature map; a third correlation between the position in the fifth intermediate positioning map and the texture feature in the third attention feature map is determined, and based on the third correlation, the texture feature in the third attention map is fused into the sixth intermediate positioning map to obtain a third texture map to present the texture of the feature; the third texture map is fused with the upsampled third positioning map by using a residual connection to obtain a third fused feature map.
[0050] In a possible implementation, the loss function is:
[0051]
[0052] in and They are BCE loss function and IoU loss respectively. GT is a binary truth map with a value range of (0,1). is the output of the i-th saliency module, i=1,2,3 Indicates upsampling.
[0053] Based on the description of the embodiment of the intelligent product selection system that can be adjusted based on real-time analysis of data, the embodiment of the present invention further discloses an intelligent product selection method that can be adjusted based on real-time analysis of data, which is specifically applied to the above system. The intelligent product selection method that can be adjusted based on real-time analysis of data includes:
[0054] Obtaining platform product information in a recent period, wherein the platform product information includes at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume;
[0055] Predicting the demand for goods in the next cycle based on the platform goods information; determining a preset number of goods types based on the demand for goods;
[0056] Obtain N products to be selected corresponding to each product type, where N is a positive integer;
[0057] Performing quality assessment on the N selected commodities, and determining a target commodity based on the quality assessment result;
[0058] The step of performing quality determination on the N selected commodities and determining the target commodity based on the quality determination result includes:
[0059] The N commodities to be selected are in the shape of bars; a quality determination process of the N commodities to be selected is filmed by a camera to obtain a commodity filming video, wherein the quality determination process comprises: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, where M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected;
[0060] Swinging the target commodity to be selected based on a preset swing amplitude and swing angle;
[0061] Determine a first video frame and a second video frame at a first angle and a second angle respectively from the video shot of the product, wherein the first angle and the second angle are two angles corresponding to the maximum swing amplitude;
[0062] Identifying the target commodity to be selected in the first video frame and the second video frame;
[0063] Determine M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively;
[0064] A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points;
[0065] Based on the first curve and the second curve, curvatures at T first feature points and T second feature points are calculated respectively, where T first feature points and T second feature points are feature points farthest from the clamping device among the M first feature points and the M second feature points;
[0066] A first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points;
[0067] determining the hardness of the target commodity to be selected based on the first average curvature and the second average curvature;
[0068] The target commodity is determined based on the hardness of each of the N commodities to be selected.
[0069] Figure 43 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The device includes: at least one processor 301, such as a central processing unit (CPU), at least one memory 302, and at least one bus 303.
[0070] The memory 302 may store program instructions, and the processor 301 may be used to call program instructions to execute an adjustable intelligent product selection system based on real-time analysis of data.
[0071] A person skilled in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electrically-erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), a solid state disk (SSD) or other optical disk storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.
[0072] It should be noted that all steps of the embodiments of the present application are performed under legal and compliant conditions, that is, all steps of the embodiments of the present application are performed with authorization.
[0073] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.
Claims
1. An intelligent product selection system that can be adjusted based on real-time analysis of data, characterized in that: The system comprises: Data acquisition module: used to acquire platform product information in the most recent period, wherein the platform product information includes at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume; Data processing module: used to predict the demand for goods in the next cycle based on the platform goods information; determine a preset number of goods types based on the demand for goods; The data acquisition module is further used to acquire N commodities to be selected corresponding to each commodity type, where N is a positive integer; A quality determination module, used to determine the quality of the N selected commodities and determine the target commodity based on the quality determination result; The step of performing quality determination on the N selected commodities and determining the target commodity based on the quality determination result includes: The N commodities to be selected are in the shape of bars; a quality determination process of the N commodities to be selected is filmed by a camera to obtain a commodity filming video, wherein the quality determination process comprises: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, where M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; Swinging the target commodity to be selected based on a preset swing amplitude and swing angle; Determine a first video frame and a second video frame at a first angle and a second angle respectively from the video shot of the product, wherein the first angle and the second angle are two angles corresponding to the maximum swing amplitude; Identifying the target commodity to be selected in the first video frame and the second video frame; Determine M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; Based on the first curve and the second curve, curvatures at T first feature points and T second feature points are calculated respectively, where the T first feature points and the T second feature points are the T feature points farthest from the clamping device among the M first feature points and the M second feature points; A first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; determining the hardness of the target commodity to be selected based on the first average curvature and the second average curvature; The target commodity is determined based on the hardness of each of the N commodities to be selected.
2. The system according to claim 1, characterized in that The determining the target commodity based on the hardness of each of the N commodities to be selected comprises: Inputting the image of each of the N commodities to be selected into the trained target neural network, and outputting the quality grade of each of the N commodities to be selected; Obtaining a score for each of the N commodities to be selected based on the hardness and the quality grade of each commodity to be selected; The target product is determined based on the score of each of the N products to be selected.
3. The system according to claim 2, characterized in that The target neural network comprises: Encoder, first feature fusion module, second feature fusion module, third feature fusion module, decoder; The encoder includes five feature extraction modules, which extract the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map in order from large to small scales; and input the first initial feature map, the second initial feature map, the third initial feature map, the fourth initial feature map, and the fifth initial feature map into the corresponding convolution layer compression channel to obtain the corresponding first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map; wherein the first feature map contains the bottom texture feature, and the fifth feature map contains the high-level semantic feature; Inputting the first feature map, the second feature map and the fifth feature map into a first feature fusion module to obtain a first fused feature map; Inputting the first feature map, the third feature map and the fifth feature map into a second feature fusion module to obtain a second fused feature map; Inputting the first feature map, the fourth feature map and the fifth feature map into a third feature fusion module to obtain a third fused feature map; The decoder comprises a first saliency module, a second saliency module and a third saliency module; The third fused feature map is input into the third saliency module to obtain a first intermediate feature map; the second fused feature map is fused with the first intermediate feature map and then input into the second saliency module to obtain a second intermediate feature map; the first fused feature map is fused with the second intermediate feature map and then input into the first saliency module to obtain a saliency map; Based on the first intermediate feature map, the second intermediate feature map, the saliency map and the true value map, supervised training is performed based on the target loss function to obtain a trained target neural network.
4. The system according to claim 3, characterized in that The step of inputting the first feature map, the second feature map and the fifth feature map into a first feature fusion module to obtain a first fused feature map comprises: the first feature fusion module comprises a first positioning module and a first texture module; The first positioning module reduces the dimension of the fifth feature map and the second feature map by average pooling, and then fuses them, and then inputs two fully connected layers with activation functions to generate a first channel attention map, and then uses the first channel attention map and the second feature map to perform channel attention calculation to obtain a first positioning map; In the first texture module, Upsampling the first positioning map and using a self-attention mechanism to generate a first intermediate positioning map and a second intermediate positioning map; downsampling the first feature map and using a self-attention mechanism to generate a first attention feature map, wherein the upsampled first positioning map has the same dimension as the downsampled first feature map; Determine a first correlation between a position in the first intermediate positioning map and a texture feature in the first attention feature map, and fuse the texture feature in the first attention map into the second intermediate positioning map based on the first correlation to obtain a first texture map to present a texture of the feature; The first texture map is fused with the upsampled first positioning map using a residual connection to obtain a first fused feature map.
5. The system according to claim 4, characterized in that The step of inputting the first feature map, the third feature map, and the fifth feature map into a second feature fusion module to obtain a second fused feature map includes: The second feature fusion module includes a second positioning module and a second texture module; The second positioning module reduces the dimension of the fifth feature map and the third feature map by average pooling, and then fuses them, and then inputs two fully connected layers with activation functions to generate a second channel attention map, and then uses the second channel attention map and the third feature map to perform channel attention calculation to obtain a second positioning map; In the second texture module, Upsampling the second positioning map and using a self-attention mechanism to generate a third intermediate positioning map and a fourth intermediate positioning map; downsampling the first feature map and using a self-attention mechanism to generate a second attention feature map, wherein the upsampled second positioning map has the same dimension as the downsampled first feature map; Determine a second correlation between the position in the third intermediate positioning map and the texture feature in the second attention feature map, and fuse the texture feature in the second attention map into the fourth intermediate positioning map based on the second correlation to obtain a second texture map to present the texture of the feature; The second texture map is fused with the upsampled second positioning map using a residual connection to obtain a second fused feature map.
6. The system according to claim 5, characterized in that The step of inputting the first feature map, the fourth feature map, and the fifth feature map into a third feature fusion module to obtain a third fused feature map comprises: The third feature fusion module includes a third positioning module and a third texture module; The third positioning module reduces the dimension of the fifth feature map and the fourth feature map by average pooling, and then fuses them, and then inputs two fully connected layers with activation functions to generate a third channel attention map, and then uses the third channel attention map and the fourth feature map to perform channel attention calculation to obtain a third positioning map; In the third texture module, Upsampling the third positioning map and using a self-attention mechanism to generate a fifth intermediate positioning map and a sixth intermediate positioning map; downsampling the first feature map and using a self-attention mechanism to generate a third attention feature map, wherein the upsampled third positioning map has the same dimension as the downsampled first feature map; Determine a third correlation between the position in the fifth intermediate positioning map and the texture feature in the third attention feature map, and fuse the texture feature in the third attention map into the sixth intermediate positioning map based on the third correlation to obtain a third texture map to present the texture of the feature; The third texture map is fused with the upsampled third positioning map by using a residual connection to obtain a third fused feature map.
7. The system according to claim 6, characterized in that The loss function is: in and They are BCE loss function and IoU loss respectively. GT is a binary truth map with a value range of (0,1). is the output of the i-th saliency module, i=1,2,3 Indicates upsampling.
8. An intelligent product selection method that can be adjusted based on real-time analysis of data, characterized in that: The method comprises: Obtaining platform product information in a recent period, wherein the platform product information includes at least one of the following: product click volume, product purchase volume, product add-on volume, and product search volume; Predicting the demand for goods in the next cycle based on the platform goods information; determining a preset number of goods types based on the demand for goods; Obtain N products to be selected corresponding to each product type, where N is a positive integer; Performing quality assessment on the N selected commodities, and determining a target commodity based on the quality assessment result; The step of performing quality determination on the N selected commodities and determining the target commodity based on the quality determination result includes: The N commodities to be selected are in the shape of bars; a quality determination process of the N commodities to be selected is filmed by a camera to obtain a commodity filming video, wherein the quality determination process comprises: clamping one end of the target commodity to be selected by a clamping device; determining M feature points of the target commodity to be selected, where M≥2; the M feature points are equidistant from each other and the M feature points include two endpoints closest to and farthest from the clamping device, and the target commodity to be selected is one of the N commodities to be selected; Swinging the target commodity to be selected based on a preset swing amplitude and swing angle; Determine a first video frame and a second video frame at a first angle and a second angle respectively from the video shot of the product, wherein the first angle and the second angle are two angles corresponding to the maximum swing amplitude; Identifying the target commodity to be selected in the first video frame and the second video frame; Determine M first feature points and M second feature points corresponding to the M feature points in the first video frame and the second video frame respectively; A first curve is obtained by fitting based on the M first feature points, and a second curve is obtained by fitting based on the M second feature points; Based on the first curve and the second curve, curvatures at T first feature points and T second feature points are calculated respectively, where the T first feature points and the T second feature points are the T feature points farthest from the clamping device among the M first feature points and the M second feature points; A first average curvature and a second average curvature are calculated based on the curvatures at the T first feature points and the T second feature points; determining the hardness of the target commodity to be selected based on the first average curvature and the second average curvature; The target commodity is determined based on the hardness of each of the N commodities to be selected.
9. An electronic device, characterized in that: include: a processor adapted to implement one or more instructions; as well as, A computer storage medium storing one or more instructions, wherein the one or more instructions are suitable for being loaded by the processor and executed by the adjustable intelligent product selection method based on real-time analysis of data as described in claim 8.
10. A computer storage medium, characterized in that: The computer storage medium stores one or more instructions, and the one or more instructions are suitable for being loaded by a processor and executed as claimed in claim 8, an adjustable intelligent product selection method based on real-time analysis of data.
Citation Information
Patent Citations
Hardness testing method and apparatus thereof
CN104297500A
Product quality prediction method
CN110288199A