A marketing-based image big data analysis method

Abstract

The application discloses a kind of based on marketing's image big data analysis method, comprising the following steps: S1, obtain multi-source image data and carry out pretreatment;S2, extract local area feature and match in different marketing image local area, screen corresponding visual area;S3, local area feature is configured as visual node, establishes connection edge according to spatial combination relationship, constructs visual element association graph;S4, mode search is carried out to connection structure, screen high-order visual combination structure, and repeat combination structure is merged;S5, local area is matched with visual node and locates difference visual node and difference connection edge, generates image analysis result;S6, obtain feedback data, and increment correction is carried out to visual element association graph.The application can realize the automatic analysis of marketing image, improve the accuracy of marketing image big data analysis.

Description

Technical Field

[0001] This invention relates to the field of image data analysis technology, and in particular to an image big data analysis method based on marketing. Background Technology

[0002] With the development of digital media and online promotion methods, a large amount of image data, such as posters, product promotional images, and event display images, has been generated in the marketing process. These images are usually composed of multiple visual areas, including text, main content, background, and decorative elements. Different images may also have similar content and different layout variations. Therefore, image analysis technology is needed to organize and identify a large number of marketing images.

[0003] Existing technologies typically analyze marketing images using methods such as image classification, object detection, or overall similarity comparison. These methods can identify parts of an image or find images with similar overall visual styles. However, the differences in marketing images often lie not only in the content of individual areas but also in the combination and positional relationships between multiple areas. For images with similar main content but different local layouts, existing methods struggle to make detailed distinctions and identify recurring typical visual combinations from a large number of images.

[0004] Meanwhile, when new marketing images differ from existing images only in the form of partial content replacement, missing areas, or arrangement changes, existing analysis methods can usually only provide an overall judgment of similarity or difference, making it difficult to further locate the specific visual areas that have changed and their relationships, thus affecting the accuracy of image content classification and difference analysis.

[0005] Therefore, how to provide a marketing-based image big data analysis method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an image big data analysis method based on marketing. This invention fully utilizes multi-scale feature coding, visual element association graphs, frequent subgraph mining, and graph editing distance analysis techniques. It describes in detail the processing method for parsing corresponding visual regions, mining high-order visual combination structures, and locating structural differences in images to be analyzed from a large number of marketing images. It has the advantages of accurate visual combination recognition, clear local difference location, and strong adaptability to image analysis.

[0007] According to an embodiment of the present invention, a marketing-based image big data analysis method includes the following steps: S1. Acquire multi-source image data in marketing scenarios and preprocess it to form a marketing image sample set; S2. Based on the marketing image sample set, a multi-scale feature coding method is used to extract local region features, and according to the local region features and regional distribution relationship, the local regions in different marketing images are matched and the corresponding visual regions are screened. S3. Configure local area features as visual nodes, establish connection edges based on the spatial combination relationship between visual nodes in the same marketing image, and configure the connection edge weights according to the co-occurrence frequency of spatial combination relationship in the marketing image sample set to construct a visual element association graph. S4. The frequent subgraph mining method is used to perform pattern search on the connection structure between visual nodes. Based on the connection edge weight, the high-order visual combination structure that meets the support conditions is screened out, and the graph isomorphism matching method is used to merge the repeated combination structure. S5. Match the local regions in the marketing image to be analyzed with visual nodes. Use the graph editing distance method to compare the spatial combination relationship between local regions with the higher-order visual combination structure. Locate the difference visual nodes and difference connection edges along the minimum editing path to generate image analysis results. S6. Obtain feedback data of image analysis results and incrementally correct the visual element association map using the near-end gradient update method.

[0008] Optionally, the preprocessing specifically includes: The system analyzes the image format, channel arrangement, and size parameters of multi-source image data, converts images of different formats into image matrices with a unified channel order, and performs proportional scaling and edge filling according to preset sizes. Local gradient calculation is performed on the image matrix after size processing. Based on the local gradient response distribution, images with blurred regions or missing effective regions are screened out to obtain candidate image samples. Calculate the perceptual hash value of the candidate image samples, and identify and remove duplicate images based on the Hamming distance between the perceptual hash values, while retaining non-duplicate candidate image samples; Non-duplicate candidate image samples are labeled and configured, and sample indexes are established according to image source and acquisition order to form a marketing image sample set.

[0009] Optionally, S2 specifically includes: S21. Input each marketing image in the marketing image sample set into a multi-scale feature coding network, perform hierarchical coding of image features at different scales, and perform size alignment and channel fusion on the coding results of each scale to obtain a fused feature map, wherein: The multi-scale feature coding network includes four consecutive convolutional coding layers and one feature fusion layer; The four consecutive convolutional coding layers correspond to the original scale, the 2x downsampling scale, the 4x downsampling scale, and the 8x downsampling scale, respectively. The feature fusion layer is used to perform size alignment and channel fusion on the encoding results at each scale. The marketing image is input into a continuous convolutional coding layer. A 3×3 convolutional kernel is used to perform layer-by-layer convolutional coding on the marketing image. The image coding results at four scales are extracted according to downsampling ratios of 1x, 2x, 4x and 8x respectively. Perform a 1×1 convolution transformation on the image encoding results at each of the four scales to uniformly adjust the channel dimension of the image encoding results at each scale to 256 dimensions; Using the image encoding result corresponding to the original scale as the size alignment benchmark, upsampling is performed on the image encoding results corresponding to the two-times downsampling scale, the four-times downsampling scale, and the eight-times downsampling scale respectively to obtain four aligned encoding results with consistent spatial size. The four aligned encoding results are concatenated along the channel direction, and a 3×3 convolution operation is used to compress the channel dimension of the concatenated result to 256 dimensions to obtain the fused feature map. S22. Select local regions based on the feature response peaks in the fused feature map, perform pooling on the feature codes corresponding to each local region to obtain local region features, and record the center position and area ratio of each local region in the corresponding marketing image. S23. For any two marketing images, establish candidate matching pairs between their local regions, and configure the region matching cost based on the feature distance, center position offset and area ratio difference between the local region features in the candidate matching pairs. S24. Using the optimal transmission matching method, the candidate matching pairs are transmitted and allocated according to the regional matching cost, and the local regions that meet the preset matching conditions are selected as the corresponding visual regions.

[0010] Optionally, S24 specifically includes: S241. For any two marketing images, determine the transmission quality share of the source region and the target region based on the area ratio of each local region, and construct the regional transmission cost relationship based on the corresponding regional matching cost of the candidate matching. S242. Introduce entropy regularization constraints into the regional transmission cost relationship, and use the Sinkhorn iterative method to solve the transmission coupling between the source region and the target region, so that the solution results satisfy the transmission quality share constraints of the source region and the target region. S243. Based on the transmission coupling amount, analyze the degree of concentration of the distribution of the target region corresponding to the same source region, and filter out candidate matching pairs with dispersed transmission coupling amounts that do not meet the main correspondence condition. S244. For the retained candidate matching pairs, jointly calculate the region matching cost and transmission coupling amount, and screen out the local region pairs that meet the preset matching conditions as the corresponding visual regions.

[0011] Optionally, S3 specifically includes: S31. Based on the matching relationship between the local features of the corresponding visual regions, the corresponding visual regions that belong to the same matching result are merged into visual nodes, and the regional position of each visual node in the corresponding marketing image is retained. S32. Within the same marketing image, analyze the combined positions of different visual nodes and determine the spatial combination relationship between visual nodes based on the center position of the area, the distance between areas, and the direction of area arrangement. S33. In the marketing image sample set, perform co-occurrence statistics on visual nodes with the same visual node combination and the same spatial combination relationship, and establish connection edges between visual nodes whose co-occurrence frequency meets the preset connection conditions. S34. Configure the weights of the connecting edges according to their co-occurrence frequency, and combine the visual nodes, connecting edges and connecting edge weights to form a visual element association graph.

[0012] Optionally, S4 specifically includes: S41. Traverse the connection edges in the visual element association graph, filter out the connection edges whose weights do not meet the preset search conditions, and configure the remaining connection edges and the visual nodes they connect to as the pattern search range. S42. The gSpan frequent subgraph mining method is adopted. Within the pattern search range, the subgraph is expanded according to the spatial combination relationship of visual nodes and connecting edges. The minimum depth-first search code is generated for the connection structure formed by the expansion, and the non-repeating candidate combination structure is retained according to the minimum depth-first search code. S43. Map the candidate combination structures to the marketing image sample set, count the number of marketing image samples containing complete candidate combination structures, and screen out candidate combination structures that meet the preset support conditions based on the statistical results to form a high-order visual combination structure, specifically including: According to the connection order of the connecting edges in the candidate combination structure, the visual node corresponding to the first connecting edge is found in the marketing image sample set, and the local regions that satisfy the spatial combination relationship within the same marketing image are matched. Based on the matched visual nodes, the matching continues along the subsequent connecting edges to match local areas within the same marketing image that satisfy the corresponding spatial combination relationship. The total number of complete occurrences is counted when all visual nodes and connecting edges in the candidate combination structure are matched. When any connecting edge cannot be matched, the structural expansion of the current marketing image is terminated; The support of the combined structure is calculated based on the number of complete occurrences, and the candidate combined structures that meet the preset support conditions are screened as higher-order visual combined structures. S44. Using the graph isomorphic matching method, in higher-order visual combination structures, the correspondence between visual nodes and the spatial combination relationship between connecting edges are compared, and higher-order visual combination structures with isomorphic relationships are merged.

[0013] Optionally, S42 specifically includes: S421. Extract the connecting edges and the visual nodes at both ends of the connecting edges from the pattern search range, and perform depth-first search encoding according to the spatial combination relationship between the visual node identifiers and the connecting edges to form an initial subgraph. S422. Determine the rightmost path based on the initial subgraph. Perform backward expansion from the rightmost visual node to the existing visual nodes in the rightmost path, and perform forward expansion from the visual nodes in the rightmost path to the adjacent newly added visual nodes. Connect the expanded visual nodes and their corresponding spatial combination relationships to the current subgraph to form an expanded subgraph. S423. Generate a depth-first search code based on the visual node traversal order, edge extension direction, and spatial combination relationship of the edge in the extended subgraph, and compare the lexicographical order of the depth-first search codes corresponding to the same extended subgraph to determine the minimum depth-first search code. S424. Based on the depth-first search code of the extended subgraph, verify it with the minimum depth-first search code. Cut off the extended branches that do not satisfy the minimum depth-first search code, and continue to expand the rightmost path of the extended subgraph that satisfies the minimum depth-first search code to form a candidate combination structure.

[0014] Optionally, S5 specifically includes: S51. Using a multi-scale feature coding method, extract local region features in the marketing image to be analyzed, calculate the feature distance between each local region feature and the corresponding local region feature of the visual node, and establish a correspondence between the local regions whose feature distances meet the preset matching conditions and the visual nodes. S52. Based on the location of the corresponding local regions in the marketing image to be analyzed, analyze the spatial combination relationship between the local regions, and combine the corresponding visual nodes and spatial combination relationships into the combination structure to be analyzed. S53. Using the graph editing distance method, the visual nodes and connecting edges in the combined structure to be analyzed and the higher-order visual combined structure are edited and matched to obtain the editing path and the cumulative path cost. S54. Screen out the high-order visual combination structure and its editing path with the minimum cumulative path cost, locate the difference visual nodes and difference connecting edges based on the visual node editing operations and connecting edge editing operations in the editing path, and generate image analysis results.

[0015] Optionally, S53 specifically includes: S531. Traverse the visual nodes in the combined structure to be analyzed, and compare them with the visual nodes in the higher-order visual combined structure in terms of local region features. Select the corresponding nodes in order of feature distance from small to large. S532. For the selected corresponding nodes, perform node replacement editing based on the differences in local region features; for visual nodes for which no corresponding nodes are selected, perform node addition editing or node deletion editing according to the structure in which they are located. S533. Search for connecting edges along the connection relationships between the selected corresponding nodes, perform edge replacement editing on connecting edges with different spatial combination relationships, and perform edge addition or edge deletion editing on connecting edges that only exist in one side of the structure. S534. Following the execution order of node editing and edge editing, the editing path is gradually expanded, and the current editing cost is accumulated after each expansion to obtain the editing path and cumulative path cost between the combined structure to be analyzed and each higher-order visual combined structure.

[0016] Optionally, S6 specifically includes: S61. Analyze the feedback data of the image analysis results. Based on the difference confirmation and correction results in the feedback data, locate the connection edge to be corrected and its corresponding higher-order visual combination structure along the minimum editing path, where: The feedback data is a data record formed after reviewing the difference visual nodes and difference connecting edges in the image analysis results, including difference visual node confirmation information, difference connecting edge confirmation information, and correction information corresponding to mislocated difference visual nodes or difference connecting edges. S62. Based on the discrepancy between the difference confirmation results and the image analysis results, calculate the edge weight deviation of the connection edge to be corrected and the support deviation of the higher-order visual combination structure, and construct the weight update loss. S63. The proximal gradient update method is used to iteratively solve the weight update loss. In each iteration, gradient updates are performed on the connection edge weights and support weights, and the weight changes of non-corresponding feedback biases are suppressed by proximal contraction. S64. Update the connection edge weights after iterative solution to the visual element association graph, and update the support weights after iterative solution to the corresponding higher-order visual combination structure to complete the incremental correction.

[0017] The beneficial effects of this invention are: First, this invention analyzes the visual content and regional distribution relationships in marketing images through multi-scale feature encoding and local region matching, which can identify corresponding visual regions between different images and reduce the problem of insufficient recognition of local layout differences in overall similarity analysis.

[0018] Secondly, this invention constructs a visual element association graph based on the spatial combination relationship between visual regions, and extracts high-order visual combination structures by combining frequent subgraph mining and graph isomorphism matching. This enables the screening of recurring visual organization forms from a large number of marketing images, enhancing the hierarchy of image structure analysis.

[0019] Finally, this invention employs the graph editing distance method to perform structural comparison of the marketing image to be analyzed, locates the difference visual nodes and difference connecting edges along the minimum editing path, and performs incremental correction on the relevant weights, which can refine the identification of local area changes and arrangement changes, and improve the accuracy and adaptability of image analysis results. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an image big data analysis method based on marketing proposed in this invention; Figure 2 This is a flowchart illustrating the construction of a visual element association diagram for an image big data analysis method based on marketing proposed in this invention. Figure 3 This is a schematic diagram comparing the high-order visual composition structure and the minimum editing path of an image big data analysis method based on marketing proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A marketing-based image big data analysis method includes the following steps: S1. Acquire multi-source image data in marketing scenarios and preprocess it to form a marketing image sample set; S2. Based on the marketing image sample set, a multi-scale feature coding method is used to extract local region features, and according to the local region features and regional distribution relationship, the local regions in different marketing images are matched and the corresponding visual regions are screened. S3. Configure local area features as visual nodes, establish connection edges based on the spatial combination relationship between visual nodes in the same marketing image, and configure the connection edge weights according to the co-occurrence frequency of spatial combination relationship in the marketing image sample set to construct a visual element association graph. S4. The frequent subgraph mining method is used to perform pattern search on the connection structure between visual nodes. Based on the connection edge weight, the high-order visual combination structure that meets the support conditions is screened out, and the graph isomorphism matching method is used to merge the repeated combination structure. S5. Match the local regions in the marketing image to be analyzed with visual nodes. Use the graph editing distance method to compare the spatial combination relationship between local regions with the higher-order visual combination structure. Locate the difference visual nodes and difference connection edges along the minimum editing path to generate image analysis results. S6. Obtain feedback data of image analysis results and incrementally correct the visual element association map using the near-end gradient update method.

[0023] In this embodiment, the preprocessing specifically includes: The system analyzes the image format, channel arrangement, and size parameters of multi-source image data, converts images of different formats into image matrices with a unified channel order, and performs proportional scaling and edge filling according to preset sizes. Local gradient calculation is performed on the image matrix after size processing. Based on the local gradient response distribution, images with blurred regions or missing effective regions are screened out to obtain candidate image samples. Calculate the perceptual hash value of the candidate image samples, and identify and remove duplicate images based on the Hamming distance between the perceptual hash values, while retaining non-duplicate candidate image samples; Non-duplicate candidate image samples are labeled and configured, and sample indexes are established according to image source and acquisition order to form a marketing image sample set; In this embodiment, the acquired marketing images are uniformly converted into an RGB three-channel image matrix, and the pixel values ​​of each channel are mapped to the range of 0 to 1; The image is scaled proportionally to 512×512 pixels, and the edge areas not covered after scaling are filled with adjacent boundary pixels. For the image matrix after size processing, the grayscale component is extracted, and the local gradient response is calculated using the 3×3 Sobel operator to divide the image into several 16×16 pixel blocks. When the number of regions with a mean gradient response value lower than 0.02 accounts for more than 40% of the total number of regions, the corresponding image will be filtered out as a blurred image. Calculate a 64-bit perceptual hash value for the retained images. If the Hamming distance between the corresponding perceptual hash values ​​of two images is not greater than 5, they are determined to be duplicate images, and the first image sample that appears is retained according to the acquisition order. A unique number is assigned to each non-duplicate candidate image sample according to the image source identifier and the collection order, forming a marketing image sample set.

[0024] In this embodiment, S2 specifically includes: S21. Input each marketing image in the marketing image sample set into a multi-scale feature coding network, perform hierarchical coding of image features at different scales, and perform size alignment and channel fusion on the coding results of each scale to obtain a fused feature map, wherein: The multi-scale feature coding network consists of four consecutive convolutional coding layers and one feature fusion layer; The four consecutive convolutional coding layers correspond to the original scale, the 2x downsampling scale, the 4x downsampling scale, and the 8x downsampling scale, respectively. The feature fusion layer is used to perform size alignment and channel fusion on the encoding results at each scale; The marketing image is input into a continuous convolutional coding layer. A 3×3 convolutional kernel is used to perform layer-by-layer convolutional coding on the marketing image. The image coding results at four scales are extracted according to downsampling ratios of 1x, 2x, 4x and 8x respectively. Perform a 1×1 convolution transformation on the image encoding results at each of the four scales to uniformly adjust the channel dimension of the image encoding results at each scale to 256 dimensions; Using the image encoding result corresponding to the original scale as the size alignment benchmark, upsampling is performed on the image encoding results corresponding to the two-times downsampling scale, the four-times downsampling scale, and the eight-times downsampling scale respectively to obtain four aligned encoding results with consistent spatial size. The four aligned encoding results are concatenated along the channel direction, and a 3×3 convolution operation is used to compress the channel dimension of the concatenated result to 256 dimensions to obtain the fused feature map. S22. Select local regions based on the feature response peaks in the fused feature map, perform pooling on the feature codes corresponding to each local region to obtain local region features, and record the center position and area ratio of each local region in the corresponding marketing image. S23. For any two marketing images, establish candidate matching pairs between their local regions, and configure the region matching cost based on the feature distance, center position offset and area ratio difference between the local region features in the candidate matching pairs. S24. Using the optimal transmission matching method, the candidate matching pairs are transmitted and allocated according to the regional matching cost, and the local regions that meet the preset matching conditions are selected as the corresponding visual regions.

[0025] In this embodiment, S24 specifically includes: S241. For any two marketing images, determine the transmission quality share of the source region and the target region based on the area ratio of each local region, and construct the regional transmission cost relationship based on the corresponding regional matching cost of the candidate matching. S242. Introduce entropy regularization constraints into the regional transmission cost relationship, and use the Sinkhorn iterative method to solve the transmission coupling between the source region and the target region, so that the solution results satisfy the transmission quality share constraints of the source region and the target region. S243. Based on the transmission coupling amount, analyze the degree of concentration of the distribution of the target region corresponding to the same source region, and filter out candidate matching pairs with dispersed transmission coupling amounts that do not meet the main correspondence condition. S244. For the retained candidate matching pairs, jointly calculate the region matching cost and transmission coupling amount, and screen out the local region pairs that meet the preset matching conditions as the corresponding visual regions. In this embodiment, for any two marketing images, the area ratio of each local region is used as the transmission quality share, and the transmission relationship between regions is established based on the region matching cost. The Sinkhorn iterative method with entropy regularization constraint is used to solve the transmission coupling. The entropy regularization parameter is set to 0.05, the maximum number of iterations is set to 100, and the iteration stops when the change in coupling between two adjacent iterations is less than 0.001. For the same source region, when the maximum coupling allocation ratio of the corresponding target region is not less than 0.60 and the region matching cost is not greater than 0.30, the source region and the target region are screened as the corresponding visual regions.

[0026] In this embodiment, S3 specifically includes: S31. Based on the matching relationship between the local region features of the corresponding visual regions, the corresponding visual regions belonging to the same matching result are merged into visual nodes, and the regional position of each visual node in the corresponding marketing image is retained, specifically including: Extract the local region features, the corresponding marketing image identifier, and the region location of the corresponding visual region, and configure the corresponding visual regions with matching relationships as units to be merged; The unjoined-set merging method is adopted to perform connectivity retrieval on the matching relationships between the units to be merged, and merge the units to be merged with direct matching relationships or continuous matching relationships into a candidate merge set; The local region features within the candidate merge set are aggregated, and the feature distance between each local region feature and the aggregated feature is calculated. The corresponding visual regions whose feature distances meet the preset merge conditions are retained in the same candidate merge set to form visual nodes. Establish a correspondence between visual nodes and their corresponding visual regions, and retain the marketing image identifier and region location of each corresponding visual region in the correspondence; S32. Within the same marketing image, analyze the combined positions of different visual nodes and determine the spatial combination relationship between visual nodes based on the center position of the area, the distance between areas, and the direction of area arrangement. S33. In the marketing image sample set, perform co-occurrence statistics on visual nodes with the same visual node combination and the same spatial combination relationship, and establish connection edges between visual nodes whose co-occurrence frequency meets the preset connection conditions. S34. Configure the weights of the connecting edges according to their co-occurrence frequency, and combine the visual nodes, connecting edges and connecting edge weights to form a visual element association graph.

[0027] In this embodiment, S4 specifically includes: S41. Traverse the connection edges in the visual element association graph, filter out the connection edges whose weights do not meet the preset search conditions, and configure the remaining connection edges and the visual nodes they connect to as the pattern search range. S42. The gSpan frequent subgraph mining method is adopted. Within the pattern search range, the subgraph is expanded according to the spatial combination relationship of visual nodes and connecting edges. The minimum depth-first search code is generated for the connection structure formed by the expansion, and the non-repeating candidate combination structure is retained according to the minimum depth-first search code. S43. Map the candidate combination structures to the marketing image sample set, count the number of marketing image samples containing complete candidate combination structures, and screen out candidate combination structures that meet the preset support conditions based on the statistical results to form a high-order visual combination structure, specifically including: According to the connection order of the connecting edges in the candidate combination structure, the visual node corresponding to the first connecting edge is found in the marketing image sample set, and the local regions that satisfy the spatial combination relationship within the same marketing image are matched. Based on the matched visual nodes, the matching continues along the subsequent connecting edges to match local areas within the same marketing image that satisfy the corresponding spatial combination relationship. The total number of complete occurrences is counted when all visual nodes and connecting edges in the candidate combination structure are matched. When any connecting edge cannot be matched, the structural expansion of the current marketing image is terminated; The support of the combined structure is calculated based on the number of complete occurrences, and the candidate combined structures that meet the preset support conditions are screened as higher-order visual combined structures. S44. Using the graph isomorphic matching method, in higher-order visual combination structures, the correspondence between visual nodes and the spatial combination relationship between connecting edges are compared, and higher-order visual combination structures with isomorphic relationships are merged.

[0028] In this embodiment, S42 specifically includes: S421. Extract the connecting edges and the visual nodes at both ends of the connecting edges from the pattern search range, and perform depth-first search encoding according to the spatial combination relationship between the visual node identifiers and the connecting edges to form an initial subgraph. S422. Determine the rightmost path based on the initial subgraph. Perform backward expansion from the rightmost visual node to the existing visual nodes in the rightmost path, and perform forward expansion from the visual nodes in the rightmost path to the adjacent newly added visual nodes. Connect the expanded visual nodes and their corresponding spatial combination relationships to the current subgraph to form an expanded subgraph. S423. Generate a depth-first search code based on the visual node traversal order, edge extension direction, and spatial combination relationship of the edge in the extended subgraph, and compare the lexicographical order of the depth-first search codes corresponding to the same extended subgraph to determine the minimum depth-first search code. S424. Based on the depth-first search code of the extended subgraph, verify it with the minimum depth-first search code. Cut off the extended branches that do not satisfy the minimum depth-first search code, and continue to expand the rightmost path of the extended subgraph that satisfies the minimum depth-first search code to form a candidate combination structure.

[0029] In this embodiment, S5 specifically includes: S51. Using a multi-scale feature coding method, extract local region features in the marketing image to be analyzed, calculate the feature distance between each local region feature and the corresponding local region feature of the visual node, and establish a correspondence between the local regions whose feature distances meet the preset matching conditions and the visual nodes. S52. Based on the location of the corresponding local regions in the marketing image to be analyzed, analyze the spatial combination relationship between the local regions, and combine the corresponding visual nodes and spatial combination relationships into the combination structure to be analyzed. S53. Using the graph editing distance method, the visual nodes and connecting edges in the combined structure to be analyzed and the higher-order visual combined structure are edited and matched to obtain the editing path and the cumulative path cost. S54. Screen out the high-order visual combination structure and its editing path with the minimum cumulative path cost, locate the difference visual nodes and difference connecting edges based on the visual node editing operations and connecting edge editing operations in the editing path, and generate image analysis results.

[0030] In this embodiment, S53 specifically includes: S531. Traverse the visual nodes in the combined structure to be analyzed, and compare them with the visual nodes in the higher-order visual combined structure in terms of local region features. Select the corresponding nodes in order of feature distance from small to large. S532. For the selected corresponding nodes, perform node replacement editing based on the differences in local region features; for visual nodes for which no corresponding nodes are selected, perform node addition editing or node deletion editing according to the structure in which they are located. S533. Search for connecting edges along the connection relationships between the selected corresponding nodes, perform edge replacement editing on connecting edges with different spatial combination relationships, and perform edge addition or edge deletion editing on connecting edges that only exist in one side of the structure. S534. Following the execution order of node editing and edge editing, the editing path is gradually expanded, and the current editing cost is accumulated after each expansion to obtain the editing path and cumulative path cost between the combined structure to be analyzed and each higher-order visual combined structure.

[0031] In this embodiment, S6 specifically includes: S61. Analyze the feedback data of the image analysis results. Based on the difference confirmation and correction results in the feedback data, locate the connection edge to be corrected and its corresponding higher-order visual combination structure along the minimum editing path, where: The feedback data consists of data records generated after reviewing the differential visual nodes and differential connecting edges in the image analysis results. These records include confirmation information for differential visual nodes, confirmation information for differential connecting edges, and correction information for mislocated differential visual nodes or differential connecting edges. S62. Based on the discrepancy between the difference confirmation results and the image analysis results, calculate the edge weight deviation of the connection edge to be corrected and the support deviation of the higher-order visual combination structure, and construct the weight update loss. S63. The proximal gradient update method is used to iteratively solve the weight update loss. In each iteration, gradient updates are performed on the connection edge weights and support weights, and the weight changes of non-corresponding feedback biases are suppressed by proximal contraction. S64. Update the connection edge weights after iterative solution to the visual element association graph, and update the support weights after iterative solution to the corresponding higher-order visual combination structure to complete the incremental correction.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an online marketing image material analysis scenario. This scenario accumulated a large number of images, including product display images, promotional images, and page illustrations. Images of the same type often share similar main content but differ in text areas, decorative areas, or layout relationships. Conventional whole-image similarity analysis methods tend to group images with different local structures into the same category, making it difficult to further pinpoint the specific visually changing areas.

[0033] In this scenario, 2680 marketing images were selected as the raw data. These images were uniformly converted to RGB three-channel format and adjusted to 512×512 pixels. Blurry images were removed using local gradient response, and duplicate images were removed using 64-bit perceptual hash values, ultimately retaining 2416 images to form a marketing image sample set. Subsequently, a multi-scale feature encoding network was used to process each marketing image, fusing the encoding results from different scales into a 256-dimensional fusion feature map, from which local region features were extracted.

[0034] For local regions in different images, matching is performed by combining region features, center position and area ratio. The Sinkhorn iterative method is used to solve the transmission coupling between regions, where the entropy regularization parameter is set to 0.05 and the maximum number of iterations is set to 100, and a total of 18,642 corresponding visual regions are screened.

[0035] After merging the corresponding visual regions, 327 visual nodes are formed. Connection edges are established based on the spatial combination relationships of each visual node within the same marketing image, and the weights of these edges are assigned according to the co-occurrence frequency of the spatial combination relationships, resulting in a visual element association graph containing 1489 connection edges. A frequent subgraph mining method is used to perform a structural search on the visual element association graph, and repeated combination structures are processed through graph isomorphism matching to obtain 46 higher-order visual combination structures, which are used to represent typical visual layout forms that recur in marketing images.

[0036] Another 320 marketing images not involved in the composition were selected for analysis, of which 214 were confirmed to have local area replacements, missing parts, or arrangement changes. The graph edit distance method was used to compare the spatial composition relationships in the marketing images to be analyzed with higher-order visual composition structures, locating the differing visual nodes and differing connecting edges along the minimum edit path.

[0037] The initial analysis correctly identified 201 images with discrepancies, achieving a recognition rate of 93.93%. Of the 396 actual visual nodes with discrepancies, 371 were correctly located, and of the 283 actual connecting edges with discrepancies, 260 were correctly located. Further analysis, based on feedback data from the review, led to incremental adjustments to the support weights of the connecting edge weights and the higher-order visual combination structure. This resulted in the correct identification of 207 images with discrepancies, increasing the recognition rate to 96.73%. The number of incorrectly located visual nodes with discrepancies decreased from 42 to 25, and the number of incorrectly located connecting edges with discrepancies decreased from 37 to 21.

[0038] Table 1. Validation Results of Visual Structure Difference Analysis for Marketing Images

[0039] As can be seen from the data in Table 1, the present invention can extract recurring high-order visual combination structures from a large number of marketing images and locate local regional changes and spatial combination changes in the image to be analyzed; after incremental correction, the accuracy of image structure difference analysis is further improved.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A marketing-based image big data analysis method, characterized in that, Includes the following steps: S1. Acquire multi-source image data in marketing scenarios and preprocess it to form a marketing image sample set; S2. Based on the marketing image sample set, a multi-scale feature coding method is used to extract local region features, and according to the local region features and regional distribution relationship, the local regions in different marketing images are matched and the corresponding visual regions are screened. S3. Configure local area features as visual nodes, establish connection edges based on the spatial combination relationship between visual nodes in the same marketing image, and configure the connection edge weights according to the co-occurrence frequency of spatial combination relationship in the marketing image sample set to construct a visual element association graph. S4. The frequent subgraph mining method is used to perform pattern search on the connection structure between visual nodes. Based on the connection edge weight, the high-order visual combination structure that meets the support conditions is screened out, and the graph isomorphism matching method is used to merge the repeated combination structure. S5. Match the local regions in the marketing image to be analyzed with visual nodes. Use the graph editing distance method to compare the spatial combination relationship between local regions with the higher-order visual combination structure. Locate the difference visual nodes and difference connection edges along the minimum editing path to generate image analysis results. S6. Obtain feedback data of image analysis results and incrementally correct the visual element association map using the near-end gradient update method.

2. The image big data analysis method based on marketing as described in claim 1, characterized in that, The preprocessing specifically includes: The system analyzes the image format, channel arrangement, and size parameters of multi-source image data, converts images of different formats into image matrices with a unified channel order, and performs proportional scaling and edge filling according to preset sizes. Local gradient calculation is performed on the image matrix after size processing. Based on the local gradient response distribution, images with blurred regions or missing effective regions are screened out to obtain candidate image samples. Calculate the perceptual hash value of the candidate image samples, and identify and remove duplicate images based on the Hamming distance between the perceptual hash values, while retaining non-duplicate candidate image samples; Non-duplicate candidate image samples are labeled and configured, and sample indexes are established according to image source and acquisition order to form a marketing image sample set.

3. The image big data analysis method based on marketing according to claim 1, characterized in that, S2 specifically includes: S21. Input each marketing image in the marketing image sample set into a multi-scale feature coding network, perform hierarchical coding of image features at different scales, and perform size alignment and channel fusion on the coding results of each scale to obtain a fused feature map, wherein: The multi-scale feature coding network includes four consecutive convolutional coding layers and one feature fusion layer; The four consecutive convolutional coding layers correspond to the original scale, the 2x downsampling scale, the 4x downsampling scale, and the 8x downsampling scale, respectively. The feature fusion layer is used to perform size alignment and channel fusion on the encoding results at each scale. The marketing image is input into a continuous convolutional coding layer. A 3×3 convolutional kernel is used to perform layer-by-layer convolutional coding on the marketing image. The image coding results at four scales are extracted according to downsampling ratios of 1x, 2x, 4x and 8x respectively. Perform a 1×1 convolution transformation on the image encoding results at each of the four scales to uniformly adjust the channel dimension of the image encoding results at each scale to 256 dimensions; Using the image encoding result corresponding to the original scale as the size alignment benchmark, upsampling is performed on the image encoding results corresponding to the two-times downsampling scale, the four-times downsampling scale, and the eight-times downsampling scale respectively to obtain four aligned encoding results with consistent spatial size. The four aligned encoding results are concatenated along the channel direction, and a 3×3 convolution operation is used to compress the channel dimension of the concatenated result to 256 dimensions to obtain the fused feature map. S22. Select local regions based on the feature response peaks in the fused feature map, perform pooling on the feature codes corresponding to each local region to obtain local region features, and record the center position and area ratio of each local region in the corresponding marketing image. S23. For any two marketing images, establish candidate matching pairs between their local regions, and configure the region matching cost based on the feature distance, center position offset and area ratio difference between the local region features in the candidate matching pairs. S24. Using the optimal transmission matching method, the candidate matching pairs are transmitted and allocated according to the regional matching cost, and the local regions that meet the preset matching conditions are selected as the corresponding visual regions.

4. The image big data analysis method based on marketing according to claim 3, characterized in that, S24 specifically includes: S241. For any two marketing images, determine the transmission quality share of the source region and the target region based on the area ratio of each local region, and construct the regional transmission cost relationship based on the corresponding regional matching cost of the candidate matching. S242. Introduce entropy regularization constraints into the regional transmission cost relationship, and use the Sinkhorn iterative method to solve the transmission coupling between the source region and the target region, so that the solution results satisfy the transmission quality share constraints of the source region and the target region. S243. Based on the transmission coupling amount, analyze the degree of concentration of the distribution of the target region corresponding to the same source region, and filter out candidate matching pairs with dispersed transmission coupling amounts that do not meet the main correspondence condition. S244. For the retained candidate matching pairs, jointly calculate the region matching cost and transmission coupling amount, and screen out the local region pairs that meet the preset matching conditions as the corresponding visual regions.

5. The image big data analysis method based on marketing according to claim 1, characterized in that, S3 specifically includes: S31. Based on the matching relationship between the local features of the corresponding visual regions, the corresponding visual regions that belong to the same matching result are merged into visual nodes, and the regional position of each visual node in the corresponding marketing image is retained. S32. Within the same marketing image, analyze the combined positions of different visual nodes and determine the spatial combination relationship between visual nodes based on the center position of the area, the distance between areas, and the direction of area arrangement. S33. In the marketing image sample set, perform co-occurrence statistics on visual nodes with the same visual node combination and the same spatial combination relationship, and establish connection edges between visual nodes whose co-occurrence frequency meets the preset connection conditions. S34. Configure the weights of the connecting edges according to their co-occurrence frequency, and combine the visual nodes, connecting edges and connecting edge weights to form a visual element association graph.

6. The image big data analysis method based on marketing according to claim 1, characterized in that, S4 specifically includes: S41. Traverse the connection edges in the visual element association graph, filter out the connection edges whose weights do not meet the preset search conditions, and configure the remaining connection edges and the visual nodes they connect to as the pattern search range. S42. The gSpan frequent subgraph mining method is adopted. Within the pattern search range, the subgraph is expanded according to the spatial combination relationship of visual nodes and connecting edges. The minimum depth-first search code is generated for the connection structure formed by the expansion, and the non-repeating candidate combination structure is retained according to the minimum depth-first search code. S43. Map the candidate combination structures to the marketing image sample set, count the number of marketing image samples containing complete candidate combination structures, and screen out candidate combination structures that meet the preset support conditions based on the statistical results to form a high-order visual combination structure, specifically including: According to the connection order of the connecting edges in the candidate combination structure, the visual node corresponding to the first connecting edge is found in the marketing image sample set, and the local regions that satisfy the spatial combination relationship within the same marketing image are matched. Based on the matched visual nodes, the matching continues along the subsequent connecting edges to match local areas within the same marketing image that satisfy the corresponding spatial combination relationship. The total number of complete occurrences is counted when all visual nodes and connecting edges in the candidate combination structure are matched. When any connecting edge cannot be matched, the structural expansion of the current marketing image is terminated; The support of the combined structure is calculated based on the number of complete occurrences, and the candidate combined structures that meet the preset support conditions are screened as higher-order visual combined structures. S44. Using the graph isomorphic matching method, in higher-order visual combination structures, the correspondence between visual nodes and the spatial combination relationship between connecting edges are compared, and higher-order visual combination structures with isomorphic relationships are merged.

7. The image big data analysis method based on marketing according to claim 6, characterized in that, S42 specifically includes: S421. Extract the connecting edges and the visual nodes at both ends of the connecting edges from the pattern search range, and perform depth-first search encoding according to the spatial combination relationship between the visual node identifiers and the connecting edges to form an initial subgraph. S422. Determine the rightmost path based on the initial subgraph. Perform backward expansion from the rightmost visual node to the existing visual nodes in the rightmost path, and perform forward expansion from the visual nodes in the rightmost path to the adjacent newly added visual nodes. Connect the expanded visual nodes and their corresponding spatial combination relationships to the current subgraph to form an expanded subgraph. S423. Generate a depth-first search code based on the visual node traversal order, edge extension direction, and spatial combination relationship of the edge in the extended subgraph, and compare the lexicographical order of the depth-first search codes corresponding to the same extended subgraph to determine the minimum depth-first search code. S424. Based on the depth-first search code of the extended subgraph, verify it with the minimum depth-first search code. Cut off the extended branches that do not satisfy the minimum depth-first search code, and continue to expand the rightmost path of the extended subgraph that satisfies the minimum depth-first search code to form a candidate combination structure.

8. The image big data analysis method based on marketing according to claim 1, characterized in that, S5 specifically includes: S51. Using a multi-scale feature coding method, extract local region features in the marketing image to be analyzed, calculate the feature distance between each local region feature and the corresponding local region feature of the visual node, and establish a correspondence between the local regions whose feature distances meet the preset matching conditions and the visual nodes. S52. Based on the location of the corresponding local regions in the marketing image to be analyzed, analyze the spatial combination relationship between the local regions, and combine the corresponding visual nodes and spatial combination relationships into the combination structure to be analyzed. S53. Using the graph editing distance method, the visual nodes and connecting edges in the combined structure to be analyzed and the higher-order visual combined structure are edited and matched to obtain the editing path and the cumulative path cost. S54. Screen out the high-order visual combination structure and its editing path with the minimum cumulative path cost, locate the difference visual nodes and difference connecting edges based on the visual node editing operations and connecting edge editing operations in the editing path, and generate image analysis results.

9. The image big data analysis method based on marketing as described in claim 8, characterized in that, S53 specifically includes: S531. Traverse the visual nodes in the combined structure to be analyzed, and compare them with the visual nodes in the higher-order visual combined structure in terms of local region features. Select the corresponding nodes in order of feature distance from small to large. S532. For the selected corresponding nodes, perform node replacement editing based on the differences in local region features; for visual nodes for which no corresponding nodes are selected, perform node addition editing or node deletion editing according to the structure in which they are located. S533. Search for connecting edges along the connection relationships between the selected corresponding nodes, perform edge replacement editing on connecting edges with different spatial combination relationships, and perform edge addition or edge deletion editing on connecting edges that only exist in one side of the structure. S534. Following the execution order of node editing and edge editing, the editing path is gradually expanded, and the current editing cost is accumulated after each expansion to obtain the editing path and cumulative path cost between the combined structure to be analyzed and each higher-order visual combined structure.

10. The image big data analysis method based on marketing according to claim 1, characterized in that, S6 specifically includes: S61. Analyze the feedback data of the image analysis results. Based on the difference confirmation and correction results in the feedback data, locate the connection edge to be corrected and its corresponding higher-order visual combination structure along the minimum editing path, where: The feedback data is a data record formed after reviewing the difference visual nodes and difference connecting edges in the image analysis results, including difference visual node confirmation information, difference connecting edge confirmation information, and correction information corresponding to mislocated difference visual nodes or difference connecting edges. S62. Based on the discrepancy between the difference confirmation results and the image analysis results, calculate the edge weight deviation of the connection edge to be corrected and the support deviation of the higher-order visual combination structure, and construct the weight update loss. S63. The proximal gradient update method is used to iteratively solve the weight update loss. In each iteration, gradient updates are performed on the connection edge weights and support weights, and the weight changes of non-corresponding feedback biases are suppressed by proximal contraction. S64. Update the connection edge weights after iterative solution to the visual element association graph, and update the support weights after iterative solution to the corresponding higher-order visual combination structure to complete the incremental correction.

Images