Multi-parameter fusion art work image display system

Abstract

The present application relates to the technical field of parameter fusion, and more particularly to a multi-parameter fusion art work image display system, comprising: a multi-parameter feature extraction module that extracts original art work image through an art work feature parameter extraction model based on an ArtBench backbone network architecture to generate multi-scale art work features; a feature recognition module that uses multiple paths of an art work feature recognition model based on an FPN architecture to generate multi-scale art work feature maps from the multi-scale art work features; a multi-dimensional evaluation parameter generation module that uses a feature fusion evaluation model based on an attention mechanism to generate composition parameters and color parameters from the multi-scale art work feature maps; and a display management module that adjusts the original art work image based on the composition parameters and color parameters for display. The present application makes the composition logic and color levels in the art work more intuitively presented, helping the audience better understand the artistic expression core of the art work.

Description

Technical Field

[0001] This invention relates to the field of parameter fusion technology, and in particular to a multi-parameter fusion system for displaying images of artworks. Background Technology

[0002] With the integration of digital technology and the art field, the digital display of artworks has become an important means to improve the efficiency of art dissemination and appreciation experience, and it also plays a crucial role in cultivating children's artistic literacy and aesthetics. Traditional art display systems mostly rely on static image presentation and resolution optimization, failing to fully integrate the artistic characteristics of artworks, such as composition, color, and texture, for intelligent processing, especially in enhancing the artistic features of artworks. This makes it difficult for viewers to intuitively perceive the deeper artistic value of the works.

[0003] Currently, in terms of recognizing artistic features in artworks, existing neural network models, such as the pre-trained ArtBench model, only output high-dimensional, abstract image features. These image features lack artistic semantic interpretability, and users cannot directly determine which part of the artwork is good / bad or how the colors are matched based on the image features.

[0004] Furthermore, the Cornerstone Assessment Model (MCA), as the theoretical basis for research on quality aesthetic education ability assessment products, is often used in primary school art teaching evaluation and teaching practice activities, but it cannot be directly used to evaluate the artistic characteristics of artworks.

[0005] Therefore, how to accurately identify artistic features through multi-parameter fusion based on the ArtBench model and the Cornerstone Assessment Model (MCA), and how to enhance the display of the composition and color highlights of artworks in accordance with professional art knowledge, are technical problems that need to be solved. Summary of the Invention

[0006] To this end, the present invention provides a multi-parameter fusion art image display system. By extracting features at multiple scales and dynamically enhancing the artistic feature extraction capabilities based on the ArtBench model and the professional evaluation parameters of the MCA model, abstract artistic features are transformed into visual enhancement effects. This makes the compositional logic and color hierarchy in the artwork more intuitive, lowers the understanding threshold for non-professional viewers, helps viewers better understand the core artistic expression of the artwork, and assists in cultivating viewers' art appreciation ability.

[0007] To achieve the above objectives, this invention proposes a multi-parameter fusion system for displaying artwork images, comprising:

[0008] The multi-parameter feature extraction module is used to extract multi-scale art features from the original artwork image through an artwork feature parameter extraction model based on the ArtBench backbone network architecture.

[0009] The feature recognition module is used to generate a multi-scale artwork feature map by passing the multi-scale artwork features through multiple paths of the artwork feature recognition model based on the FPN architecture.

[0010] The multi-dimensional evaluation parameter generation module is used to generate composition parameters and color parameters by using the feature fusion evaluation model based on attention mechanism and MCA cornerstone evaluation model to generate the feature map of the multi-scale artwork.

[0011] The display management module is used to adjust the composition and color of the original artwork image based on the composition parameters and the color parameters to generate an enhanced artwork image, and to display the original artwork image and the enhanced artwork image.

[0012] The first RevCol network, the second RevCol network, and the MobileViT network of the artwork feature parameter extraction model are connected to the first FPN layer, the second FPN layer, and the third FPN layer of the artwork feature recognition model, respectively. The first FPN layer, the second FPN layer, and the third FPN layer process the texture features, local image features, and global image features of multi-scale artwork features in sequence.

[0013] Furthermore, the multi-parameter feature extraction module includes:

[0014] A texture extraction unit is used to extract texture features from the original artwork image through a first RevCol network to generate the texture features;

[0015] A local feature extraction unit is used to extract local composition color features from the texture features through a second RevCol network to generate the local image features;

[0016] The full-image composition extraction unit is used to extract global composition color features from the local image features through the third RevCol network and the MobileViT network to generate the global image features;

[0017] The convolutional receptive fields of the first RevCol network, the second RevCol network, and the third RevCol network increase sequentially.

[0018] Furthermore, the first RevCol network includes a first convolutional layer, a first RevCol convolutional layer, and a Gabor layer, and the texture extraction unit includes:

[0019] The first convolutional extraction unit is used to pre-extract features from the original artwork image through the first convolutional layer to generate convolutional texture features;

[0020] The texture orientation extraction subunit is used to analyze the orientation characteristics of the original artwork image through a Gabor layer to generate texture orientation features;

[0021] The texture feature reconstruction subunit is used to pass the original artwork image through the first RevCol convolutional layer to generate reconstructed texture features;

[0022] The feature dynamic stitching subunit is used to weight and stitch the convolutional texture features, the texture direction features, and the reconstructed texture features based on dynamic weights of the original artwork image to generate the texture features.

[0023] Furthermore, the local feature extraction unit includes:

[0024] The RevCol convolutional collaborative extraction subunit is used to extract and concatenate the texture features through the second convolutional layer and the second RevCol convolutional layer of the second RevCol network to generate local fusion features.

[0025] The spatial attention fusion subunit is used to generate local attention-enhanced features by passing the local fusion features through a spatial attention mechanism based on Sobel level kernels;

[0026] The golden mask fusion subunit is used to weight the local attention enhancement features based on the golden section mask to generate the local image features.

[0027] Furthermore, the full-image composition extraction unit includes:

[0028] The convolutional full-image extraction subunit is used to pass the local image features through the dilated convolutional layer of the third RevCol network to generate color variation features;

[0029] The RevCol full-image color composition extraction subunit is used to encode the color change features through the third RevCol convolutional layer, and then pass them through the RevCol color component decoder and RevCol color component decoder respectively to generate full-image color components and full-image composition components.

[0030] The color composition attention subunit is used to generate color features by passing the color change features and the full image color components through the color attention of the MobileViT network, and to generate composition features by passing the color change features and the full image composition components through the composition attention of the MobileViT network.

[0031] The dynamic fusion subunit is used to fuse the full-image color components, the full-image composition components, the color features, and the composition features based on dynamic weighting coefficients to generate the global image features.

[0032] In particular, a collaborative architecture of a three-level RevCol network and a MobileViT network is adopted. By gradually increasing the receptive field and number of channels of the convolution, texture features, local image features and global composition color features are extracted in sequence to achieve full-dimensional feature coverage from micro-texture to macro-composition. Through texture direction analysis of Gabor layer, spatial attention mechanism of Sobel horizontal kernel and weighted optimization of golden section mask, attention to key areas is improved, and aesthetically sensitive areas in artworks are enhanced.

[0033] Furthermore, the multi-parameter feature extraction module also includes:

[0034] The model training and optimization unit is used to construct a global style difference loss function based on the labeled image genre of the samples and the KL divergence of the global image features, and to train and optimize the feature parameter extraction model of the artwork through the global style difference loss function.

[0035] Furthermore, the artwork feature recognition model is built on the BiFPN architecture, and the multi-scale artwork feature map includes a brushstroke direction map, a visual emphasis map, and a style emphasis map;

[0036] The feature recognition module is used to generate a style emphasis map by passing the global image features through the third FPN layer, to generate a visual emphasis map by passing the style emphasis map and the local image features through the second FPN layer and weightedly fusing them with the style emphasis map, and to generate a brushstroke direction map by passing the visual emphasis map and the texture features through the first FPN layer and weightedly fusing them with the visual emphasis map.

[0037] Furthermore, the multi-dimensional evaluation parameter generation module includes:

[0038] A spatial channel attention unit is used to generate a fused feature map by channel spatial collaborative attention of the feature maps of the multi-scale artworks.

[0039] The art dimension evaluation unit is used to calculate the visual centroid coordinates, golden ratio deviation, symmetry index, color emotion mapping, and color harmony based on the fused feature map.

[0040] The composition and color evaluation unit is used to generate the composition parameters and color parameters by applying the golden ratio deviation and symmetry index, the color emotion mapping and color harmony to a scoring framework based on the MCA cornerstone evaluation model.

[0041] Furthermore, the display management module includes:

[0042] The information box display unit is used to annotate the information box on the original artwork image based on the visual center coordinates to generate a first enhanced artwork image;

[0043] A transparency adjustment unit is used to adjust the transparency of the original artwork image based on the color harmony to generate an information box annotation map of the second enhanced artwork image.

[0044] Furthermore, the artwork image display system also includes:

[0045] The model collaborative training module constructs an orthogonal constraint loss function based on the spatial attention matrix and channel attention vector of the channel spatial collaborative attention, and trains the artwork feature recognition model and the feature fusion evaluation model through the orthogonal constraint loss function.

[0046] In particular, by employing BiFPN bidirectional cross-scale connectivity and weighted feature fusion mechanism, the generated brushstroke direction map, visual emphasis map, and style emphasis map realize the hierarchical expression of artistic features. The information box annotation function is based on the visual center coordinate fusion of the hierarchically expressed brushstroke direction and visual emphasis to help the audience understand the composition logic. The transparency is dynamically adjusted according to the color harmony to optimize the image layers and help the audience understand the color logic.

[0047] Compared with the prior art, the beneficial effects of the present invention are that, through multi-scale feature extraction and dynamic enhancement of the art feature extraction capability based on the ArtBench model and the professional evaluation parameters of the MCA model, the present invention transforms abstract art features into visual enhancement effects, making the compositional logic and color hierarchy in artworks more intuitive, lowering the understanding threshold for non-professional viewers, helping viewers better understand the core artistic expression of artworks, and thus assisting in cultivating viewers' art appreciation ability.

[0048] In particular, this invention adopts a collaborative architecture of a three-level RevCol network and a MobileViT network. By gradually increasing the receptive field and number of channels of the convolution, it sequentially extracts texture features, local image features, and global composition color features, achieving full-dimensional feature coverage from micro-texture to macro-composition. Through texture direction analysis of Gabor layers, spatial attention mechanism of Sobel horizontal kernels, and weighted optimization of golden section masks, it enhances attention to key areas and focuses on enhancing aesthetically sensitive areas in artworks.

[0049] In particular, this invention employs a BiFPN bidirectional cross-scale connection and weighted feature fusion mechanism to generate brushstroke direction maps, visual emphasis maps, and style emphasis maps, which realize the hierarchical expression of artistic features. The information box annotation function is based on the visual center coordinate fusion of the hierarchically expressed brushstroke direction and visual emphasis to help viewers understand the composition logic. The transparency is dynamically adjusted according to the color harmony to optimize the image layers and help viewers understand the color logic. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the structure of the art image display system according to an embodiment of the present invention;

[0051] Figure 2 This is a flowchart illustrating the multi-dimensional evaluation parameter generation module of the art artwork image display system according to an embodiment of the present invention.

[0052] Figure 3 This is a flowchart illustrating the artwork feature recognition model of the artwork image display system according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the overall model architecture of the art image display system according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0055] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0056] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0057] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0058] like Figures 1 to 4 As shown, this invention provides a multi-parameter fusion art image display system. By using multi-scale feature extraction and dynamic enhancement based on the ArtBench model's art feature extraction capabilities and the MCA model's professional evaluation parameters, abstract art features are transformed into visual enhancement effects. This makes the compositional logic and color layers in the artwork more intuitively presented, lowers the understanding threshold for non-professional viewers, helps viewers better understand the core artistic expression of the artwork, and assists in cultivating viewers' art appreciation abilities.

[0059] like Figures 1 to 4 As shown, this embodiment proposes a multi-parameter fusion art image display system, including:

[0060] The multi-parameter feature extraction module is used to extract multi-scale art features from the original artwork image through an artwork feature parameter extraction model based on the ArtBench backbone network architecture.

[0061] The feature recognition module is used to generate a multi-scale artwork feature map by passing the multi-scale artwork features through multiple paths of the artwork feature recognition model based on the FPN architecture.

[0062] The multi-dimensional evaluation parameter generation module is used to generate composition parameters and color parameters by using the feature fusion evaluation model based on attention mechanism and MCA cornerstone evaluation model to generate the feature map of the multi-scale artwork.

[0063] The display management module is used to adjust the composition and color of the original artwork image based on the composition parameters and the color parameters to generate an enhanced artwork image, and to display the original artwork image and the enhanced artwork image.

[0064] The first RevCol network, the second RevCol network, and the MobileViT network of the artwork feature parameter extraction model are connected to the first FPN layer, the second FPN layer, and the third FPN layer of the artwork feature recognition model, respectively. The first FPN layer, the second FPN layer, and the third FPN layer process the texture features, local image features, and global image features of multi-scale artwork features in sequence.

[0065] Preferably, the artwork feature recognition model based on the FPN architecture is a bidirectional feature fusion (BiFPN) model. In a bottom-up path, this model upsamples and fuses features from high-resolution texture features to low-resolution global image features, propagating semantic information of the features to generate feature maps (labels). In a top-down path, the model fuses features from low-resolution global image features to high-resolution texture features, incorporating detailed information. Furthermore, weighted bidirectional connections are used at each fusion step to achieve more comprehensive information exchange.

[0066] Preferably, the artwork feature recognition model can also be a conventional FPN architecture with only a bottom-up path. The process of the artwork feature recognition model of the FPN architecture is as follows: the bottom-up feature map is passed through a 3x3 convolution kernel of the feature pyramid network to obtain C1, and then upsampled to obtain C2. C2 is passed through a 1x1 convolution kernel to obtain C3. C1, C2 and C3 are added together in a skip manner to eliminate the aliasing effect caused by upsampling. After at least two repetitions, a multi-scale artwork feature map is obtained.

[0067] It is understandable that the features of multi-scale artworks include texture features, local image features, and global image features, which are respectively generated by the artwork feature recognition model as brushstroke direction map, visual emphasis map, and style emphasis map of multi-scale artwork feature map.

[0068] Among them, texture features are high-resolution features generated by the convolution of the first RevCol network corresponding to small scales, and brushstroke direction map is the technique that reflects the microscopic level of brushstroke direction. For example, in analyzing Van Gogh's "Starry Night": the density of brushstrokes with consistent direction in the swirling brushstrokes is high, so the consistency of direction is low, while the brushstrokes in the village area are relatively straight and the consistency of direction is strong. The brushstroke direction map can reflect that the swirling brushstrokes have a strong emotional expression, while the village area has a peaceful emotional expression.

[0069] The local image features correspond to the medium-resolution features generated by the convolution of the second RevCol network at a medium scale. The visual focus map conforms to the distribution of visual appeal reflected by the compositional balance principle in the artwork. The compositional balance principle is preferably the golden ratio. For example, the visual focus of "Starry Night" is on the houses of the village near the center of the picture, and the visual focus of "Mona Lisa" is on the mouth of the figure.

[0070] The global image features correspond to the low-resolution features generated by the convolution of the third RevCol network at a low scale. The style emphasis map is a saliency map that marks the style discriminative region, revealing the essential characteristics of different art movements. For example, if the art feature recognition model determines that the light and shadow boundary of the global image features is similar to the Impressionist sample, then the style emphasis map is a saliency map of the light and shadow boundary. If the art feature recognition model determines that the geometric edges of the global image features are similar to the Cubist sample, then the style emphasis map is a saliency map of the geometric edges.

[0071] In summary, in particular, by employing an art feature parameter extraction model with an improved pre-trained ArtBench backbone network architecture and an art feature recognition model with an FPN architecture, we can collaboratively analyze the brushstroke direction, visual emphasis, and stylistic emphasis of artworks. This allows the enhanced artwork images displayed to better assist viewers in quickly understanding the emotional expression of brushstrokes, the emphasis of visual composition, and the overall style of artworks. This lowers the understanding threshold for non-professional viewers, helps them better understand the core artistic expression of artworks, and thus helps cultivate their art appreciation abilities.

[0072] Furthermore, such as Figure 2 and Figure 4 As shown, the multi-parameter feature extraction module includes:

[0073] A texture extraction unit is used to extract texture features from the original artwork image through a first RevCol network to generate the texture features;

[0074] A local feature extraction unit is used to extract local composition color features from the texture features through a second RevCol network to generate the local image features;

[0075] The full-image composition extraction unit is used to extract global composition color features from the local image features through the third RevCol network and the MobileViT network to generate the global image features;

[0076] The convolutional receptive fields of the first RevCol network, the second RevCol network, and the third RevCol network increase sequentially, while the convolutional kernel size decreases sequentially.

[0077] In particular, texture features are shallow detail features of the artwork image, which are small-scale brushstroke direction features; local image features are mid-level features, including local elements of the artwork image, such as objects with clear meanings in the artwork, including the moon, villages, fruits, etc.; global image features are high-level features, including the composition of the entire image, such as vortex composition, partitioned composition, straight edge composition, etc.

[0078] Specifically, the first convolutional layer (Conv1_x) of the first RevCol network has a receptive field of 35×35, the second convolutional layer (Conv2_x) of the second RevCol network has a receptive field of 131×131, and the dilated convolution of the third RevCol network has a receptive field of 987×987.

[0079] like Figure 4 As shown, the first RevCol network further includes a first convolutional layer, a first RevCol convolutional layer, and a Gabor layer, and the texture extraction unit includes:

[0080] The first convolutional extraction unit is used to pre-extract features from the original artwork image through the first convolutional layer to generate convolutional texture features;

[0081] The texture orientation extraction subunit is used to analyze the orientation characteristics of the original artwork image through a Gabor layer to generate texture orientation features;

[0082] The texture feature reconstruction subunit is used to pass the original artwork image through the first RevCol convolutional layer to generate reconstructed texture features;

[0083] The feature dynamic stitching subunit is used to weight and stitch the convolutional texture features, the texture direction features, and the reconstructed texture features based on dynamic weights of the original artwork image to generate the texture features.

[0084] In particular, the Gabor layer employs a multi-directional filter bank to accurately extract the direction of the brush strokes.

[0085] Specifically, the process of generating texture features can be represented as:

[0086] ,

[0087] ,

[0088] ,

[0089] ,

[0090] ,

[0091] In the formula, For convolutional texture features, express Activation function For batch normalization, This represents a 7×7 convolution kernel. Represents images of original artworks. The Gabor function represents the original artwork image and the Gabor filter. The convolution result, Indicates the direction of the Gabor filter. Indicates texture direction features, They represent, Indicates the reconstruction of texture features, Let G represent the decoding function of the first RevCol convolutional layer, and let F represent the two-path convolutional block transformation functions of the first RevCol convolutional layer (RevCol Column), respectively. This represents the output of the first path of the first RevCol convolutional layer, which is then used as the input of the second path. , These represent the feature maps input to the current path, and are used to divide the original artwork image into regions to generate... , , This represents element-wise multiplication (Hadamard product), used to highlight texture details through a mask. Represents the convolution result of the Gabor filter. The texture mask matrix, Represents texture features, This indicates that the convolutional texture features, the texture direction features, and the reconstructed texture features are weighted and concatenated, where , There are two dynamic weights. ,in This represents the result of the Sobel gradient operator calculation on the original artwork image. This represents the Frobenius norm of the result, used to measure the overall strength of the image gradient. This represents the average pooling operation, used to reduce dimensionality and extract the average color features of the entire image.

[0092] like Figure 4 As shown, the local feature extraction unit further includes:

[0093] The RevCol convolutional collaborative extraction subunit is used to extract and concatenate the texture features through the second convolutional layer and the second RevCol convolutional layer of the second RevCol network to generate local fusion features.

[0094] The spatial attention fusion subunit is used to generate local attention-enhanced features by passing the local fusion features through a spatial attention mechanism based on Sobel level kernels;

[0095] The golden mask fusion subunit is used to weight the local attention enhancement features based on the golden section mask to generate the local image features.

[0096] In particular, deep fusion of texture features is achieved through the dual-layer convolutional structure of the second RevCol network.

[0097] like Figure 4 As shown, the full-image composition extraction unit further includes:

[0098] The convolutional full-image extraction subunit is used to pass the local image features through the dilated convolutional layer of the third RevCol network to generate color variation features;

[0099] The RevCol full-image color composition extraction subunit is used to encode the color change features through the third RevCol convolutional layer, and then pass them through the RevCol color component decoder and RevCol color component decoder respectively to generate full-image color components and full-image composition components.

[0100] The color composition attention subunit is used to generate color features by passing the color change features and the full image color components through the color attention of the MobileViT network, and to generate composition features by passing the color change features and the full image composition components through the composition attention of the MobileViT network.

[0101] The dynamic fusion subunit is used to fuse the full-image color components, the full-image composition components, the color features, and the composition features based on dynamic weighting coefficients to generate the global image features.

[0102] In particular, a collaborative architecture of a three-level RevCol network and a MobileViT network is adopted. By gradually increasing the receptive field and number of channels of the convolution, texture features, local image features and global composition color features are extracted in sequence to achieve full-dimensional feature coverage from micro-texture to macro-composition. Through texture direction analysis of Gabor layer, spatial attention mechanism of Sobel horizontal kernel and weighted optimization of golden section mask, attention to key areas is improved, and aesthetically sensitive areas in artworks are enhanced.

[0103] Specifically, the process of generating local image features can be represented as:

[0104] ,

[0105] ·

[0106] ,

[0107] ,

[0108] In the formula, For local fusion features, Indicates texture features The second convolutional layer that performs the convolution operation. This represents the operation of the second RevCol convolutional layer on the texture features. This represents the inverse reconstruction operator, which performs inverse decoding on the features to recover local structural information. This indicates a feature of enhanced local attention. This represents the SIGMOD activation function. This represents a 3×3 convolution operation. express function, This represents the Sobel horizontal kernel, i.e. , This represents the golden ratio mask, where x and y are the pixel coordinates of the original artwork image. It is the golden ratio. It represents the standard deviation and controls the smoothness and diffusion range of the mask. The higher the value, the wider the mask coverage and the better the smoothing effect. The smaller the value, the more concentrated the mask is at the center point. Representing local image features, This indicates that the local attention enhancement features at pixel x and y are normalized during the calculation. This indicates that the pixel x and y points are weighted using the golden ratio mask.

[0109] like Figure 4 As shown, the full-image composition extraction unit further includes:

[0110] The convolutional full-image extraction subunit is used to pass the local image features through the dilated convolutional layer of the third RevCol network to generate color variation features;

[0111] The RevCol full-image color composition extraction subunit is used to encode the color change features through the third RevCol convolutional layer, and then pass them through the RevCol color component decoder and RevCol color component decoder respectively to generate full-image color components and full-image composition components.

[0112] The color composition attention subunit is used to generate color features by passing the color change features and the full image color components through the color attention of the MobileViT network, and to generate composition features by passing the color change features and the full image composition components through the composition attention of the MobileViT network.

[0113] The dynamic fusion subunit is used to fuse the full-image color components, the full-image composition components, the color features, and the composition features based on dynamic weighting coefficients to generate the global image features.

[0114] In particular, by using dilated convolution, the receptive field is expanded by 2.3 times without increasing the parameters, which can capture large areas of color gradation in artworks.

[0115] Specifically, the process of generating global image features can be represented as:

[0116] ,

[0117] ,

[0118] ,

[0119] ,

[0120] ,

[0121] In the formula, Characteristic of color change, Indicates a hollow convolutional layer. This represents the SIGMOD activation function. This indicates a void ratio of 4, 8, or 12. This represents the weight matrix of the dilated convolution with a dilation rate d. Representing local image features, , These respectively represent features dominated by composition and features that combine color and composition. , This represents the feature extraction and stacking structure of the third RevCol convolutional layer. These represent color characteristic variables and compositional color characteristic variables, respectively. This represents the color feature extraction function constructed by the convolutional network. This represents the graph feature extraction function used in the construction of a convolutional network. , These represent the color components and compositional components of the entire image, respectively. This represents the RevCol color component decoder built using a deconvolutional network (DeconvNet). This represents the RevCol color component decoder constructed using a 1×1 convolution and a sigmoid activation function. , These represent the mapping token sequence and the residual join token sequence, respectively. This indicates that the color change features are segmented into image patches and linear projection maps. This represents a multi-head self-attention mechanism that captures long-range dependencies through multiple attention heads. Layer normalization (LN) normalizes the feature dimensions of each token in the mapped token sequence, ensuring stable training. , These respectively represent color characteristics and compositional characteristics. express function, This represents the attention mechanism query matrix (Query) generated based on the color components and composition components of the entire image. , These represent the key and value matrices of the attention mechanism generated based on color change features, respectively. This indicates the attention head dimension, therefore color analysis and composition analysis can be achieved through different query matrices. Represents global image features. This represents four weighting coefficients, where For dynamic weighting coefficients, This represents the luminance gradient of the original artwork image calculated for the luminance channel in the HSV color space, used to quantize edge intensity. This represents the brightness gradient, used for normalization. The information entropy calculation represents the original art image, reflecting the complexity and information content of the grayscale distribution.

[0122] Furthermore, the multi-parameter feature extraction module also includes:

[0123] The model training and optimization unit is used to construct a global style difference loss function based on the labeled image genre of the samples and the KL divergence of the global image features, and to train and optimize the feature parameter extraction model of the artwork through the global style difference loss function.

[0124] In particular, by calculating the KL divergence between the sample-labeled genres and the global image features generated by the model, a global style difference loss function is constructed, enabling the model to quantitatively learn the style distribution features of different art genres.

[0125] Specifically, the global style difference loss function can be expressed as:

[0126] ( ,

[0127] In the formula, The global style difference loss function is... This represents the Kullback-Leibler divergence calculation, which measures the difference between two probability distributions. Represents global image features The style probability of the entire image is calculated using a multilayer perceptron. This represents a sample labeled with an image genre.

[0128] like Figure 3 , 4 As shown, the artwork feature recognition model is further constructed based on the BiFPN architecture, and the multi-scale artwork feature map includes a brushstroke direction map, a visual emphasis map, and a style emphasis map;

[0129] The feature recognition module is used to generate a style emphasis map by passing the global image features through the third FPN layer, to generate a visual emphasis map by passing the style emphasis map and the local image features through the second FPN layer and weightedly fusing them with the style emphasis map, and to generate a brushstroke direction map by passing the visual emphasis map and the texture features through the first FPN layer and weightedly fusing them with the visual emphasis map.

[0130] In particular, the BiFPN architecture achieves deep interaction between global style features and local detail features through a two-way path design from top to bottom and bottom to top. The third FPN layer directly receives global image features and captures the iconic elements of the genre through high-level semantic channels. The second FPN layer performs weighted fusion to dynamically balance artistic style and local composition. The first FPN layer realizes full-link analysis from macro composition to micro brushstrokes.

[0131] Specifically, the process by which the art feature recognition model generates visual emphasis maps and brushstroke direction maps can be represented as follows:

[0132] weight ,

[0133] weight ,

[0134] In the formula, For weighted index, express function, This represents the convolution operation. This indicates a splicing operation. These represent the style emphasis map, the output of the second FPN layer or the visual emphasis map, and the output of the third FPN layer, respectively, where i and j are 3 and 2 or 2 and 1. To indicate visual emphasis or brushstroke direction. Indicates an upsampling operation. This indicates a downsampling operation.

[0135] like Figure 4 As shown, the multi-dimensional evaluation parameter generation module further includes:

[0136] A spatial channel attention unit is used to generate a fused feature map by channel spatial collaborative attention of the feature maps of the multi-scale artworks.

[0137] The art dimension evaluation unit is used to calculate the visual centroid coordinates, golden ratio deviation, symmetry index, color emotion mapping, and color harmony based on the fused feature map.

[0138] The composition and color evaluation unit is used to generate the composition parameters and color parameters by applying the golden ratio deviation and symmetry index, the color emotion mapping and color harmony to a scoring framework based on the MCA cornerstone evaluation model.

[0139] In particular, by introducing artistic theory parameters such as visual center of gravity coordinates and golden ratio deviation, the compositional balance in traditional art theory is transformed into calculable deviation values. The generated composition / color parameters can directly drive the intelligent adjustment of the display management module.

[0140] Specifically, the process of generating the composition parameters and the color parameters can be represented as follows:

[0141] ,

[0142] ,

[0143] MCA_Score ,

[0144] In the formula, For visual center of gravity coordinates, The fused feature map representing the attention output of the spatial channel is shown at a scale of SIFT gradient vector under, This represents the transpose of the pixel coordinate vector. Indicates the degree of deviation from the golden ratio. Indicates the golden ratio reference point. Represents the L2 norm. Represents the symmetry index. Indicates the width and height of the image. Indicates the dominant color, Represents the optimization function. This indicates that the condition is met. An indicator function that takes a value of 1. This represents a distance threshold. By counting the number of pixels that meet the criteria, the candidate color that covers the most pixels is selected as the dominant color of the image, reflecting the overall color tendency. Represents the emotional mapping of colors. These represent the emotion mapping matrix and the color distribution vector normalized to the dominant color, respectively. K represents the color harmony, and K represents the number of primary color guides. This represents the difference between the Lab color space coordinates of the i-th and j-th dominant colors, where MCA_Score represents the composition parameters and the color parameters. , These represent the calibration coefficients for different art styles. Based on the characteristics of Impressionism, it was set to 0.8. Based on the characteristics of the ink painting school, the value is set to 1.2. It is understandable that the deviation from the golden ratio, symmetry index, color emotional mapping, and color harmony are all evaluation criteria of the MCA cornerstone assessment model.

[0145] Furthermore, the display management module includes:

[0146] The information box display unit is used to annotate the information box on the original artwork image based on the visual center coordinates to generate a first enhanced artwork image;

[0147] A transparency adjustment unit is used to adjust the transparency of the original artwork image based on the color harmony to generate an information box annotation map of the second enhanced artwork image.

[0148] Specifically, the process of generating the first enhanced artwork image and the second enhanced artwork image can be represented as follows:

[0149] ,

[0150] ,

[0151] In the formula, This indicates the first enhanced artwork image. For visual center of gravity coordinates, Indicates the dynamic offset coefficient. , Indicates the degree of deviation from the golden ratio. This indicates the second enhanced artwork image. Indicates color harmony. This represents the color distribution vector normalized to the dominant color. The values ​​for the white space enhancement coefficient and the complexity suppression coefficient are preferably 0.6 and 0.4, respectively.

[0152] In particular, by employing BiFPN bidirectional cross-scale connectivity and weighted feature fusion mechanism, the generated brushstroke direction map, visual emphasis map, and style emphasis map realize the hierarchical expression of artistic features. The information box annotation function is based on the visual center coordinate fusion of the hierarchically expressed brushstroke direction and visual emphasis to help the audience understand the composition logic. The transparency is dynamically adjusted according to the color harmony to optimize the image layers and help the audience understand the color logic.

[0153] Furthermore, the artwork image display system also includes:

[0154] The model collaborative training module constructs an orthogonal constraint loss function based on the spatial attention matrix and channel attention vector of the channel spatial collaborative attention, and trains the artwork feature recognition model and the feature fusion evaluation model through the orthogonal constraint loss function.

[0155] In particular, the feature recognition and fusion evaluation model is jointly trained by using an orthogonal constraint loss function, which forces the orthogonality of the spatial attention matrix and the channel attention vector and reduces feature redundancy.

[0156] Specifically, the orthogonal constraint loss function can be expressed as:

[0157] ,

[0158] In the formula, This represents the orthogonal constraint loss function. , These represent the spatial attention matrix and channel attention vectors of channel-space collaborative attention, respectively.

[0159] In this embodiment, by using multi-scale feature extraction and dynamic enhancement based on the ArtBench model's art feature extraction capabilities and the MCA model's professional evaluation parameters, abstract art features are transformed into visual enhancement effects. This makes the compositional logic and color hierarchy in artworks more intuitively presented, lowering the understanding threshold for non-professional viewers and helping them better understand the core artistic expression of artworks, thereby assisting in cultivating viewers' art appreciation abilities. A three-level RevCol network and MobileViT network collaborative architecture is adopted. By gradually increasing the convolutional receptive field and the number of channels, texture features, local image features, and global compositional color features are extracted sequentially, achieving full-dimensional feature coverage from micro-texture to macro-composition. Through texture direction analysis of Gabor layers, spatial attention mechanism of Sobel horizontal kernels, and weighted optimization of the golden section mask, attention to key areas is enhanced, with a focus on strengthening aesthetically sensitive areas in artworks. By employing BiFPN bidirectional cross-scale connectivity and weighted feature fusion mechanism, the generated brushstroke direction map, visual emphasis map, and style emphasis map realize the hierarchical expression of artistic features. The information box annotation function is based on the visual center coordinate fusion of the hierarchically expressed brushstroke direction and visual emphasis to help the audience understand the composition logic. The transparency is dynamically adjusted according to the color harmony to optimize the image layers and help the audience understand the color logic.

[0160] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0161] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-parameter fusion art work image display system, characterized in that, The method comprises the following steps: A multi-parameter feature extraction module is used to extract features of an original artwork image through an ArtBench backbone network architecture-based artwork feature parameter extraction model to generate multi-scale artwork features; A feature recognition module is used to pass the multi-scale artwork features through multiple paths of an FPN architecture-based artwork feature recognition model to generate multi-scale artwork feature maps; A multi-dimensional evaluation parameter generation module is used to pass the multi-scale artwork feature maps through a feature fusion evaluation model based on an attention mechanism and an MCA cornerstone evaluation model to generate composition parameters and color parameters; A display management module is used to adjust the composition and color of the original artwork image based on the composition parameters and the color parameters to generate an enhanced artwork image, and display the original artwork image and the enhanced artwork image. The first RevCol network, the second RevCol network, and the MobileViT network of the artwork feature parameter extraction model are connected to the first FPN layer, the second FPN layer, and the third FPN layer of the artwork feature recognition model, respectively, and the first FPN layer, the second FPN layer, and the third FPN layer sequentially process the texture features, the local image features, and the global image features of the multi-scale artwork features.

2. The multi-parameter fused artwork image presentation system of claim 1, wherein, The multi-parameter feature extraction module comprises: A texture extraction unit is used to extract texture features of the original artwork image through a first RevCol network to generate the texture features; A local feature extraction unit is used to extract local composition and color features of the texture features through a second RevCol network to generate the local image features; A global composition extraction unit is used to extract global composition and color features of the local image features through a third RevCol network and a MobileViT network to generate the global image features; The convolution receptive fields of the first RevCol network, the second RevCol network, and the third RevCol network increase sequentially.

3. The multi-parameter fused artwork image presentation system of claim 2, wherein, The first RevCol network comprises a first convolution layer, a first RevCol convolution layer, and a Gabor layer, and the texture extraction unit comprises: A first convolution extraction unit is used to pre-extract features of the original artwork image through the first convolution layer to generate convolution texture features; A texture direction extraction subunit is used to analyze the direction characteristics of the original artwork image through the Gabor layer to generate texture direction features; A texture feature reconstruction subunit is used to generate reconstructed texture features of the original artwork image through the first RevCol convolution layer; A feature dynamic splicing subunit is used to weight and splice the convolution texture features, the texture direction features, and the reconstructed texture features based on dynamic weights of the original artwork image to generate the texture features.

4. The multi-parameter fused artwork image presentation system of claim 2, wherein, The local feature extraction unit comprises: The convolution RevCol cooperative extraction subunit is configured to extract and splice the texture features through a second convolution layer and a second RevCol convolution layer of a second RevCol network to generate local fusion features. The spatial attention fusion subunit is configured to generate local attention enhanced features by fusing the local fusion features through a spatial attention mechanism based on a Sobel horizontal kernel. The golden mask fusion subunit is configured to weight the local attention enhanced features based on a golden section mask to generate the local image features.

5. The multi-parameter fused artwork image presentation system of claim 2, wherein, The full-image composition extraction unit includes: The convolution full-image extraction subunit is configured to generate color change features by fusing the local image features through a dilated convolution layer of a third RevCol network. The RevCol full-image color composition extraction subunit is configured to generate full-image color components and full-image composition components by decoding the color change features through a third RevCol convolution layer and then through a RevCol color component decoder and a RevCol color component decoder, respectively. The color composition attention subunit is configured to generate color features by fusing the color change features and the full-image color components through color attention of a MobileViT network, and generate composition features by fusing the color change features and the full-image composition components through composition attention of the MobileViT network. The dynamic fusion subunit is configured to fuse the full-image color components, the full-image composition components, the color features, and the composition features based on dynamic weighting coefficients to generate the global image features.

6. The multi-parameter fused artwork image presentation system of claim 2, wherein, The multi-parameter feature extraction module further includes: The model training optimization unit is configured to construct a global style difference loss function based on a KL divergence between a labeled image genre of a sample and the global image features, and train and optimize the art work feature parameter extraction model through the global style difference loss function.

7. The multi-parameter fused artwork image presentation system according to any one of claims 1 to 6, characterized in that, The art work feature recognition model is constructed based on a BiFPN architecture, and the multi-scale art work feature map includes a brush stroke direction map, a visual focus map, and a style focus map. The feature recognition module is configured to generate a style focus map by fusing the global image features through the third FPN layer, generate a visual focus map by fusing the style focus map and the local image features through the second FPN layer and then weighting the style focus map, and generate a brush stroke direction map by fusing the visual focus map and the texture features through the first FPN layer and then weighting the visual focus map.

8. The multi-parameter fused artwork image presentation system of claim 7, wherein, The multi-dimensional evaluation parameter generation module includes: The spatial channel attention unit is configured to generate a fusion feature map by fusing the multi-scale art work feature map through channel spatial cooperative attention. The artistic dimension evaluation unit is configured to calculate a visual gravity center coordinate, a golden section deviation degree, a symmetry index, a color emotion mapping, and a color harmony degree based on the fusion feature map. A composition color evaluation unit is configured to respectively pass the golden section deviation and the symmetry index, the color emotion mapping and the color harmony through a scoring framework based on an MCA cornerstone evaluation model to generate the composition parameters and the color parameters.

9. The multi-parameter fused artwork image presentation system of claim 8, wherein, Further, the display management module comprises: An information box display unit is configured to label an information box on the original artwork image based on the visual barycenter coordinates to generate a first enhanced artwork image; A transparency adjustment unit is configured to adjust the transparency of the original artwork image based on the color harmony to generate an information box labeled image of a second enhanced artwork image.

10. The multi-parameter fused artwork image presentation system of claim 9, wherein, Further comprising: A model collaborative training module is configured to construct an orthogonal constraint loss function based on the spatial attention matrix and the channel attention vector of the channel space collaborative attention, and train the artwork feature recognition model and the feature fusion evaluation model through the orthogonal constraint loss function.

Images