Coin comprehensive intelligent scoring method and system based on multi-modal data fusion

Abstract

The invention relates to the field of currency condition analysis, in particular to a currency comprehensive intelligent scoring method and system based on multi-modal data fusion. The method comprises the following steps: collecting a spectral feature image set of a to-be-detected coin; performing spectral texture response analysis on the spectral feature image set to obtain a spectral texture data sequence; performing spectral defect analysis according to the spectral texture data sequence to obtain a spectral defect response map; performing material defect change analysis based on the spectral defect response map to generate a defect change result; and performing coin condition comprehensive analysis on a defect change result, and outputting a coin condition score. According to the invention, subjective deviation among different identification personnel is reduced, and the condition analysis efficiency and accuracy of coins are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of coin appearance analysis, and in particular to a coin comprehensive intelligent scoring method and system based on multi-modal data fusion. BACKGROUND

[0002] In the modern production and circulation management process of coins, the surface quality and material integrity of coins are directly related to the circulation safety and use value of the currency. Due to the influence of casting process, material aging, circulation wear and environmental factors, coins are prone to various defects in the use process, such as surface scratches, pits, oxidation discoloration or material cracks. These defects not only affect the appearance and circulation quality of the coins, but also may lead to misjudgment in the process of automatic identification, authentication or transaction, bringing potential risks to financial management and circulation safety. Therefore, accurately and efficiently detecting and evaluating coin defects has become an important task to protect the quality and safety of currency. Traditional coin defect detection methods mainly rely on manual visual inspection or single optical imaging technology. These methods have limitations in practical application: manual detection is low in efficiency and strong in subjectivity, and it is difficult to achieve large-scale and high-precision defect identification; single optical imaging means can usually only reflect the information of a certain scale or a certain spectral range on the surface of the coin, and it is difficult to fully capture the defect characteristics hidden in the micro texture, material change or surface optical properties. The existing methods often lack multi-dimensional analysis capability, and cannot consider the topological features, material spectral information and defect distribution patterns of the coin at the same time, which limits the accuracy and reliability of defect identification. SUMMARY

[0003] To solve the above technical problems, the present application provides a coin comprehensive intelligent scoring method and system based on multi-modal data fusion to solve at least one of the above technical problems.

[0004] To achieve the above purpose, the present application provides a coin comprehensive intelligent scoring method based on multi-modal data fusion, comprising the following steps: Step S1: collecting a set of spectral feature images of the coin to be detected; performing spectral texture response analysis on the set of spectral feature images to obtain a sequence of spectral texture data; Step S2: performing spectral defect analysis according to the sequence of spectral texture data to obtain a spectral defect response spectrum; Step S3: performing material defect change analysis based on the spectral defect response spectrum to generate a defect change result; Step S4: performing coin appearance comprehensive analysis on the defect change result to output a coin appearance score.

[0005] In the present application, a coin comprehensive intelligent scoring system based on multi-modal data fusion is provided for performing the coin comprehensive intelligent scoring method based on multi-modal data fusion as described above, comprising: An image acquisition unit is used to acquire a set of spectral feature images of the coin to be detected; and to perform spectral texture response analysis on the set of spectral feature images to obtain a spectral texture data sequence. The defect analysis unit is used to perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map. The defect prediction unit is used to analyze material defect changes based on spectral defect response maps and generate defect change results. The condition scoring unit is used to perform a comprehensive analysis of the coin's condition based on the results of defect changes, and outputs a coin condition score.

[0006] The beneficial effects of this invention are as follows: By acquiring multi-band, multi-scale spectral feature images (such as visible light, near-infrared, and short-wave infrared), the reflection, absorption, and scattering characteristics of the coin surface at different wavelengths can be obtained, thereby revealing subtle structural differences that are difficult to identify with the naked eye, providing a more comprehensive foundation of original information for subsequent defect analysis. Spectral texture response analysis combines spatial texture features with the spectral dimension, effectively amplifying and quantifying minute changes such as wear, oxidation, corrosion, and indentation on the coin surface in the spectral texture data sequence, improving the ability to identify early or latent defects. By performing defect analysis on the spectral texture data sequence, abnormal response areas can be simultaneously located in the spectral and spatial dimensions, forming a spectral defect response map, thereby accurately distinguishing normal material areas from areas with defects such as wear, rust, cracks, and contamination. The spectral defect response map integrates multi-band information, effectively avoiding misjudgments caused by single lighting conditions or differences in surface reflection, making the identification of complex defects on the coin surface more stable and reliable, especially suitable for highly reflective or ancient coins. Based on existing defect response maps, the changing trends of material defects can be predicted, simulating further wear or corrosion that coins may undergo during storage, circulation, or environmental processes. This breaks through the traditional assessment model of "only detecting the current state." Even if some defects have not yet reached a level that significantly affects the condition, the results of defect changes can identify their future deterioration risk, providing a forward-looking basis for coin maintenance, repair, or value assessment. By integrating spectral texture features, defect response intensity, defect distribution, and defect change results, a comprehensive analysis model is constructed, transforming complex multidimensional information into intuitive coin condition scores, improving the understandability and comparability of assessment results. Data-driven condition scores replace or assist human experience-based judgment, reducing subjective bias between different appraisers and making coin condition assessment more objective, fair, and repeatable. By outputting standardized coin condition scores, it is easy to use directly in applications such as automatic grading, transaction pricing, and collection management, improving the practicality and promotional value of the entire multi-dimensional analysis and assessment scheme for coin defects based on multi-scale spectral images. Attached Figure Description

[0007] Fig. 1 This is a flowchart illustrating the steps of a comprehensive intelligent scoring method for coins based on multimodal data fusion according to the present invention. Fig. 2 This is a detailed flowchart illustrating the implementation steps of step S1. Fig. 3 This is a flowchart illustrating the detailed implementation steps of step S2. Detailed Implementation

[0008] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0009] This application provides a comprehensive intelligent scoring method and system for coins based on multimodal data fusion. The executing entities of the comprehensive intelligent scoring method and system for coins based on multimodal data fusion include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, and network upload devices that can be considered as general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud-based data management system.

[0010] Please see Figs. 1 to 3 This invention provides a comprehensive intelligent scoring method for coins based on multimodal data fusion, comprising the following steps: Step S1: Collect a set of spectral feature images of the coin to be detected; perform spectral texture response analysis on the set of spectral feature images to obtain a spectral texture data sequence; Step S2: Perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map; Step S3: Analyze the material defect changes based on the spectral defect response map and generate the defect change results; Step S4: Perform a comprehensive analysis of the coin condition based on the defect changes and output the coin condition score.

[0011] In the embodiments of the present invention, see Fig. 1 This is a flowchart illustrating the steps of a comprehensive intelligent currency scoring method based on multimodal data fusion according to the present invention. In this example, the steps of the comprehensive intelligent currency scoring method based on multimodal data fusion include: Step S1: Collect a set of spectral feature images of the coin to be detected; perform spectral texture response analysis on the set of spectral feature images to obtain a spectral texture data sequence; In this embodiment, the coin to be detected is first imaged in the visible to near-infrared band (400–1000 nm), with 16–32 bands selected to balance spectral resolution and acquisition efficiency. Each band image maintains high spatial resolution (approximately 2000×2000 pixels), with each pixel corresponding to an actual size of approximately 20–30 μm, ensuring that minute scratches and localized corrosion can be captured. After acquisition, the spectral feature image set is preprocessed, including dark field correction, whiteboard correction, and spectral normalization, to eliminate the effects of sensor noise and uneven illumination. Based on the preprocessing, spectral texture response analysis is performed on each band image, using methods including multi-scale local window analysis (such as 3×3, 7×7, and 15×15 pixel windows), spectral energy distribution calculation, spectral first-order gradient mean, and inter-band correlation analysis. By integrating the spectral texture responses at different scales and bands, the spectral texture features at each spatial location are organized in spatial order to form a structured spectral texture data sequence.

[0012] Step S2: Perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map; In this embodiment, the mean, variance, and inter-band correlation of the global texture response are calculated to establish a normal response reference model. Subsequently, the feature vectors of each spatial location are compared with the reference model. When the response at multiple bands or scales deviates from the reference mean by more than a threshold (typically 2–3 times the standard deviation of the mean), the location is identified as a candidate defect region. Based on this, a spatial continuity constraint is introduced, retaining only anomalous regions that form a connected structure in space to avoid the influence of isolated noise. For each candidate defect region, spectral response intensity, texture energy distribution, and local multi-scale difference features are extracted and normalized.

[0013] Step S3: Analyze the material defect changes based on the spectral defect response map and generate the defect change results; In this embodiment, the spectral defect response map is spatially matched with material distribution information to ensure that the defect location can be accurately mapped to the corresponding material region. Subsequently, for different defect types such as oxidation, corrosion, and cracks, their response change trends under time or multi-scale conditions are extracted, including the oxidation propagation rate, corrosion depth growth rate, and crack propagation direction. The oxidation propagation rate is estimated by the diffusion rate of the spectral anomaly response per unit area, the corrosion depth growth rate is calculated by the near-infrared spectral enhancement amplitude, and the crack propagation direction is obtained by analyzing the dominant direction of the defect spatial structure in the multi-scale texture response. Combining the physical properties and historical development rates of each material, a predictive model of future defect evolution trends is constructed, resulting in quantitative defect change results.

[0014] Step S4: Perform a comprehensive analysis of the coin condition based on the defect changes and output the coin condition score.

[0015] In this embodiment, defect prediction indicators such as oxidation propagation rate, corrosion depth growth rate, and crack propagation direction are normalized to ensure consistent numerical scale. Subsequently, a weighted fusion strategy is set according to the degree of impact of defects on the coin's condition. For example, the weights for corrosion depth and crack direction are set to 0.4–0.5, and the weight for oxidation propagation rate is set to 0.3–0.4, to reflect the comprehensive impact of different defect types on visual and structural integrity. The fused indicators are mapped to a preset condition scoring range, forming a single numerical condition score. Visual information on the spatial distribution and severity of defects can be added, facilitating an intuitive understanding of the coin's preservation status.

[0016] In this embodiment, see Fig. 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Collect a set of spectral feature images of the coins to be detected; Multispectral decomposition is performed on the spectral feature image set to extract multiple independent spectral signals, including material substrate components, surface texture components, and abnormal residual signals. Spectral texture response analysis is performed based on multiple independent spectral signals to obtain spectral texture data sequences.

[0017] In this embodiment, multi-band imaging of the coin surface is performed within the visible to near-infrared wavelength range. The selected spectral range covers 400 nm to 1000 nm, and 16 to 32 discrete bands are sampled within this range at equal intervals or in an adaptive manner. During the acquisition process, the coin surface is uniformly illuminated, and multi-band image acquisition is completed under fixed geometric conditions to ensure consistent spatial correspondence between different bands. The imaging resolution is controlled at the 2000×2000 pixel level, so that the actual size of a single pixel is within the range of 20 to 30 μm, thereby reflecting features such as fine scratches, worn edges, and localized corrosion. Exposure parameters are set separately for each band to address the reflection differences in different bands, ensuring that the grayscale value distribution of each band is within the effective dynamic range. To reduce the influence of background interference and illumination inhomogeneity, dark reference and white reference correction processing is performed on the spectral response before data acquisition. Spectral data is normalized along the band dimension to ensure that the reflectance of each band is on a uniform numerical scale. The spectral curve is then smoothed with a smoothing window length of 5–9 bands to reduce the impact of random noise on the decomposition results. Subsequently, the normalized spectral feature image is represented as a spectral matrix, and a multispectral decomposition method is introduced. By combining low-rank constraints and sparsity constraints, the original spectral signal is decomposed into multiple independent components. Among them, components with smooth spectral changes and continuous spatial distribution are identified as material substrate components, used to characterize the coin's metal composition and overall surface condition; components exhibiting obvious spatial texture and stable response patterns across different bands are identified as surface texture components, used to reflect embossed patterns, relief structures, and wear characteristics; and the remaining components with lower energy, scattered spatial distribution, and spectral responses deviating from the main trend are classified as anomalous residual signals, used to characterize corrosion spots, cracks, or localized abnormal areas.

[0018] Using surface texture components as the core analysis object, a multi-scale sliding window approach is employed for local analysis in the spatial dimension. Window sizes are set to 3×3, 7×7, and 15×15 pixels to characterize fine-grained texture variations and large-scale structural features. Within each window, the spectral energy distribution, mean first-order gradient, and inter-band correlation are calculated for the corresponding spectral signal to describe the texture's response characteristics under different wavelength conditions. Anomaly residual signals are introduced to correct the texture response results. When a region exhibits anomalies in the residual signal, the texture response weight for that region is enhanced to highlight the spectral characteristics of potential defect areas. The material substrate component serves as global constraint information to eliminate the influence of differences in materials on the texture response analysis results.

[0019] In this embodiment, the specific steps for acquiring the spectral feature image set of the coin to be detected are as follows: Acquire the original spectral image of the coin to be tested; Adaptive brightness equalization is performed on the original spectral image to obtain a brightness-optimized spectral image; Spectral enhancement is performed on each band of the brightness-optimized spectral image to generate a contrast-enhanced image; Spectral consistency calibration is performed on the contrast-enhanced images to obtain a set of spectral feature images.

[0020] In this embodiment, multi-band imaging of the coin surface is performed within the visible to near-infrared wavelength range, covering a spectral range of 400 nm to 1000 nm. Within this range, 16 to 32 discrete bands are sampled to ensure a thorough depiction of different material compositions and surface conditions. A fixed geometric relationship is maintained during imaging to ensure strict spatial correspondence between the images of each band, avoiding spectral mismatch caused by viewing angle shifts. The spatial resolution is set to the level of 2000×2000 pixels, with the actual size of a single pixel controlled within the range of 20 to 30 μm to balance the needs of overall structural representation and the identification of minute defects. To address the differences in reflection intensity across different bands, exposure times are set for each band, typically controlled between 10 and 50 ms, ensuring that the grayscale value distribution of each band remains within an effective dynamic range. To reduce the influence of ambient light and background reflection, background suppression is performed during acquisition, and the original spectral response is corrected using dark and white references to eliminate brightness deviations caused by non-target factors. For each spectral band, its grayscale distribution characteristics are analyzed, and the brightness statistics of local and global regions are calculated, including average brightness, variance, and local contrast. Based on this, an adaptive brightness adjustment strategy is introduced, dynamically adjusting the gain factor according to the brightness distribution in different spatial regions to enhance areas with lower brightness while avoiding over-amplification of areas with higher brightness. The brightness equalization process is typically performed using a local window method, with the window size set to 32×32 or 64×64 pixels to balance brightness smoothness and local detail preservation. To avoid introducing noise amplification during brightness adjustment, the gain range is constrained during brightness mapping, with the upper limit generally controlled between 2.0 and 3.0.

[0021] For each spectral band, its reflection intensity variation range and texture response characteristics are analyzed, and different enhancement weights are set according to the band's sensitivity in defect identification. During enhancement, a nonlinear mapping of the spectral response amplifies the differences in low-to-medium reflectance regions while keeping high-reflectance regions relatively stable, thus avoiding detail saturation. Enhancement parameters are typically adaptively adjusted based on the band's standard deviation and energy distribution, with enhancement coefficients generally controlled within the range of 1.2 to 2.5. Simultaneously, local contrast adjustment is combined in the spatial dimension to make the texture edge regions and flat regions exhibit more obvious hierarchical changes after enhancement. Using the relative spectral morphology of each band as a constraint, the enhanced spectral response is consistently adjusted to ensure continuous and smooth response changes at the same spatial location across different bands. The response curve of each pixel in the spectral dimension is calculated and compared with the average spectral curve of its neighboring pixels. When the response of a certain band deviates from the overall trend by more than a set threshold, the response of that band is corrected. The correction threshold is typically set to a relative deviation of no more than 10% to 15% to avoid over-smoothing leading to feature loss. A global spectral reference constraint is introduced to keep the overall spectral distribution consistent with the original spectral shape, preventing the enhancement process from destroying the inherent spectral differences between different materials.

[0022] In this embodiment, the specific steps for performing spectral texture response analysis based on multiple independent spectral signals to obtain a spectral texture data sequence are as follows: Material substrate reflectance analysis is performed based on material substrate components to identify the spectral reflectance data of different materials; Frequency domain analysis of surface texture components yields spatial spectral distribution data; Local anomaly detection is performed on the abnormal residual signal to obtain abnormal spectral features; Dynamic spectral texture response analysis is performed on the abnormal spectral features, spatial spectral distribution data, and spectral reflectance data to obtain a spectral texture data sequence.

[0023] In this embodiment, for the material substrate component, the reflectance curve corresponding to each pixel is extracted in the spectral dimension and normalized to ensure that the spectral reflectance data of different regions are within a uniform scale range. Subsequently, the overall shape of the spectral reflectance curve is analyzed, focusing on the distribution and trend of reflectance intensity in different bands, such as the reflectance slope change in the visible light band (400–700 nm) and the reflectance stability in the near-infrared band (700–1000 nm). By calculating the mean, standard deviation, and inter-band ratio of reflectance in each band, a spectral reflectance feature vector for material differentiation is constructed. To reduce the influence of local noise, the material substrate component is subjected to regional aggregation processing in the spatial dimension. The aggregation window size can be set to 15×15 or 25×25 pixels. While maintaining the spatial structure of the spectral components, the surface texture component undergoes frequency domain transformation processing in the spatial dimension, mapping the spatial domain texture information to the frequency domain. By analyzing the frequency distribution, the energy distribution characteristics corresponding to different spatial frequency components are extracted. Low-frequency components mainly reflect the overall outline and large-scale structure, while mid-to-high-frequency components are closely related to fine textures, edges, and wear marks. To achieve multi-scale analysis, the frequency range is segmented and statistically analyzed in the frequency domain processing. For example, the frequency range is divided into low-frequency, mid-frequency, and high-frequency bands, and the energy proportion and directional distribution characteristics within each band are calculated separately. The analysis window size can be set to 32×32 or 64×64 pixels to maintain spatial positioning capabilities while ensuring spectral resolution.

[0024] In terms of spatial dimension, the anomalous residual signal is divided into local regions, with the segmentation window size typically set to 7×7 or 11×11 pixels to highlight local anomalous changes. For each local region, its spectral response intensity, spectral energy concentration, and spectral difference from surrounding regions are calculated. When the response of a region in multiple bands deviates from the average level of its neighborhood by more than a preset threshold, the region is marked as an anomalous candidate region. The threshold is usually determined based on the overall distribution statistics of the anomalous residual signal, and the degree of deviation can be set to 2 to 3 times the standard deviation of the mean. To avoid false detections caused by noise, spatial continuity constraints are incorporated into the anomaly detection process, retaining only anomalous regions with a certain degree of spatial connectivity.

[0025] Using the spectral reflectance data of the material substrate as a global constraint, the basic response characteristics of different regions at the material level are determined. Based on this, spatial spectral distribution data is introduced into the analysis process to characterize the variation of texture structure at different spatial frequency scales. Combined with anomalous spectral features, regions exhibiting anomalous responses are assigned higher response weights. By fusing these three types of features point-by-point along the spatial dimension, a dynamic spectral texture response vector is constructed, encompassing material information, texture structure information, and defect / anomaly information. Under multi-scale conditions, response results from different analysis window scales are combined, ensuring that the data reflects both local details and maintains overall consistency.

[0026] In this embodiment, see Fig. 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Perform spectral defect analysis based on spectral texture data sequences and label spectral defect features; By identifying differences in response characteristics of spectral defects, multiple defect types can be obtained. Adaptive spectral channel enhancement based on multiple defect types yields spectral defect response maps.

[0027] In this embodiment, a spectral texture data sequence is used as input, and the multidimensional feature vectors corresponding to each spatial location are analyzed point by point, focusing on the degree of anomaly in spectral response intensity, texture energy distribution, and multi-scale variation trends. Statistical modeling is performed on the spectral texture data sequence to calculate the mean, variance, and correlation distribution of the overall sample across each spectral channel and texture scale, constructing a normal response reference interval. Subsequently, the feature vector at each spatial location is compared with the reference interval. When its response deviates from the reference interval across multiple spectral channels or multiple texture scales, and the deviation exceeds a set threshold, the location is identified as a suspected defect area. The threshold is typically set to 2-3 times the standard deviation of the mean, based on the overall distribution characteristics, to balance the defect detection rate and the false positive rate. To improve the stability of the labeling results, a spatial continuity constraint is introduced, retaining only anomalous regions that form a connected structure in adjacent locations. For each labeled spectral defect feature, its response curve shape in the spectral dimension, energy distribution in the spatial frequency dimension, and response change pattern under multi-scale conditions are extracted. By normalizing these features, the features between different defect areas are made comparable. Subsequently, defect features were grouped and analyzed based on differences in response characteristics, focusing on the differences in response strength among different defects within specific wavelength ranges. For example, some defects exhibit high reflectivity anomalies in the visible light band, while showing little change in the near-infrared band; other defects, however, show prominent performance in high-frequency texture responses. By calculating spectral curve similarity, spectral energy distribution difference, and multi-scale response consistency index, defect features with similar response characteristics were grouped into the same type. During the classification process, a minimum inter-category difference threshold can be set; for example, features with a distance greater than a preset threshold are classified into different types to avoid mixing different defect types.

[0028] The average response intensity and response stability of each defect type in each spectral channel are statistically analyzed to determine the most discriminative combination of spectral channels for that defect type. For channels with high discriminative power, their enhancement weights are increased, with enhancement coefficients generally set in the range of 1.5–3.0. For channels with lower contribution to defect differentiation, the original response is maintained or their weights are appropriately suppressed. During enhancement, the weight changes between different channels are smoothly constrained to avoid spectral discontinuities caused by abrupt weight changes. Spatially, the enhanced spectral response is mapped back to its corresponding spatial location, making defect regions more prominent in the enhanced spectrum while maintaining relative stability in non-defect regions. By applying the above enhancement strategies to different defect types and organizing the results uniformly, a spectral defect response map that intuitively reflects the defect location, type, and response intensity is finally formed.

[0029] In this embodiment, the specific steps of step S3 are as follows: Pixel-level partitioned material recognition is performed based on the material substrate components to obtain pixel-level material recognition results; Based on the pixel-level material recognition results, region segmentation and semantic annotation are performed to obtain a material distribution semantic map; The material distribution semantic map is matched with the spectral defect response spectrum, and a material defect variation analysis is performed to generate defect variation results. These results include the oxidation propagation rate, corrosion depth growth rate, and crack propagation direction.

[0030] In this embodiment, the spectral reflectance data corresponding to each pixel in the material substrate component is used as input features. Reflectance curve morphology parameters are extracted along the spectral dimension, including reflectance values ​​for each band, inter-band reflectance ratios, and spectral slope variation characteristics. To reduce local noise interference, the spectral features are lightly smoothed while maintaining pixel resolution; the smoothing window can be set to 3×3 pixels. Subsequently, based on the similarity of spectral reflectance features, pixels are partitioned for identification, classifying pixels with similar spectral response patterns into the same material category. Material prior constraints are introduced during the identification process to ensure that the identification results conform to the typical reflectance characteristics of common coin materials in the visible and near-infrared bands. For example, pixel regions with a smooth transition in the reflectance curve around 700 nm are classified as belonging to the same material category, while regions exhibiting significant absorption characteristics in the short-wavelength band are classified as belonging to another material category. Spatially, connectivity analysis is performed on pixels of the same material category, aggregating spatially adjacent pixels with the same material category into continuous regions. A minimum region area threshold is set during aggregation, for example, not less than 100–300 pixels, to avoid interference from scattered regions caused by noise. For regions with an area smaller than a threshold, they are merged or reclassified based on the material distribution of their surrounding areas. Subsequently, the boundaries of each aggregated material region are smoothed to make the edges more continuous and avoid jagged boundaries affecting subsequent matching accuracy. Based on this, each type of material region is assigned a clear semantic label, which characterizes the material type corresponding to that region and its attribute position within the coin structure. By fusing semantic labels with spatial location information, a material distribution semantic map containing spatial coordinates, region boundaries, and material category information is generated.

[0031] The two types of spectra are aligned to ensure that the defect locations in the spectral defect response map are accurately mapped to the corresponding regions in the material distribution semantic map. Subsequently, spectral response intensity, response range, and multi-scale variation information related to defects are extracted within each material region, and defect evolution is modeled and analyzed in conjunction with material type characteristics. For oxidation defects, the oxidation propagation rate is estimated by comparing the changes in defect response intensity at different locations with spatial expansion; the rate unit can be normalized to the response growth ratio per unit area. For corrosion defects, the growth rate of corrosion depth is inferred by analyzing the enhancement trend of the spectral response in the deep sensitive band. For crack defects, the dominant direction of crack propagation is extracted by statistically analyzing the spatial extension trend of the defect response. Time scale or multi-state comparison parameters are introduced during the prediction process to enable comparative analysis of defect change trends in different material regions.

[0032] In this embodiment, the specific steps of step S4 are as follows: The morphological defect depth visual detection is performed on the original spectral image to obtain the morphological defect index; The degree of material defect is assessed based on the results of defect changes to obtain the degree of material defect. Calculate the percentage of material degradation progress and expected remaining life based on the degree of material defects; The material wear index is generated by assessing the percentage of material degradation progress and the expected remaining lifespan. A comprehensive analysis of coin condition is conducted based on the morphological defect index and the material wear index to generate a coin condition score.

[0033] In this embodiment, based on the spatial information of each band in the original spectral image, the focus is on analyzing the brightness gradient changes and edge structure responses of spectral reflectance at spatial scales. By calculating the spatial gradient of each band image, the brightness change amplitude corresponding to surface undulations, depressions, and protrusions is extracted, and the gradient consistency between multiple bands is combined to distinguish between real morphological changes and noise fluctuations. During the analysis, a multi-scale structural response analysis method is introduced, calculating the intensity of local structural changes at 3×3, 7×7, and 15×15 pixel scales, respectively, so that shallow wear and deep erosion can be distinguished at different scales. For areas with obvious depression features, the defect depth level is estimated by the attenuation amplitude of reflection intensity in adjacent bands; the larger the depth response amplitude, the higher the morphological defect level. Using predicted data such as oxidation propagation rate, corrosion depth growth rate, and crack propagation direction as input, the influence of various defects at the current time point is quantitatively analyzed. The coverage trend of oxidation defects on the material surface is judged based on the oxidation propagation rate. When the propagation rate per unit area exceeds a preset threshold, the area is determined to have entered the moderate or severe oxidation stage. The degree of damage to the internal structure of the material is assessed by evaluating the corrosion depth growth rate; areas with a high growth rate are considered to have a risk of continued material damage. For crack-type defects, the impact of the crack on the overall structural stability is assessed by analyzing the relationship between its extension direction and the main direction of the material structure.

[0034] The degree of material defects is compared with a preset material integrity reference value to calculate the proportion of the current defect level within the integrity range, thus obtaining the percentage of material degradation progress. For example, when the material defect level reaches 30% to 50% of the maximum reference value, the material can be determined to be in a moderate degradation stage. Subsequently, by combining the development rate information in the defect change results, the future degradation trend is predicted, and the further degradation range of the material at different scales is estimated through linear or piecewise growth models. The percentage of material degradation progress is regarded as a direct reflection of the current wear accumulation, while the expected remaining life is used as a reverse indicator of the material's future stability. By weighted fusion of the two, a material wear assessment model is constructed. During the fusion process, different weights are set. For example, when the assessment goal is more biased towards the current state, the weight of the percentage of degradation progress is increased; when the assessment goal emphasizes long-term preservation value, the influence ratio of the expected remaining life is increased. The weight range is usually controlled between 0.4 and 0.6 to maintain the balance of the assessment results. The morphological defect index reflects the integrity of the coin's external structure, focusing on visual factors such as surface dents, scratches, and shape distortion. The material wear index reflects the internal material condition, indicating aging, corrosion, and long-term stability. Both indices are normalized to a uniform numerical range, and then weighted according to the emphasis of the condition assessment before being merged. Generally, the weights of the morphological defect index and the material wear index are set to approximately 0.5 each to achieve a comprehensive balance between appearance and material quality. The merged result is mapped to a preset condition scoring range to generate the final coin condition score.

[0035] In this embodiment, the specific steps for performing morphological defect depth visual detection based on the original spectral image to obtain the morphological defect index are as follows: The outer contour of the coin is identified from the original spectral image, and the three-dimensional contour line is extracted. Geometric parameters of the three-dimensional contour lines are calculated to obtain the three-dimensional morphological parameters; Surface pattern analysis was performed on the original spectral image to extract the coin pattern; Perform depth visual detection of morphological defects in coin images and 3D morphological parameters, and mark the location of morphological defects; Calculate the distribution of the number of defects and the proportion of the defect area based on the location of the morphological defects; The morphological defect index is obtained by evaluating the distribution of defect quantity and the proportion of defect area.

[0036] In this embodiment, key bands of the original spectral image are selected, typically the bands with the highest reflectance contrast in visible light, for edge enhancement processing. For example, gradient filtering or the Laplacian operator is used to enhance the boundary difference between the coin and the background. Subsequently, edge detection methods are used to identify the outer contour pixels. Common methods include the Canny operator or the Sobel operator, combined with a threshold setting to control the edge response sensitivity. The threshold can be set within a range of 0.1 to 0.3 based on the pixel grayscale distribution. After edge pixel identification, a sub-pixel-level fitting method is used to fit the contour pixels into a smooth curve, and hyperspectral information is collected along the normal direction of the contour curve and converted into height information to generate a three-dimensional contour line. Based on the coordinate information of the three-dimensional contour line in the X, Y, and Z spatial dimensions, basic geometric parameters such as contour length, diameter, thickness, and curvature distribution are calculated. Curvature calculation can be performed at each sampling point of the contour curve; the curvature value is used to characterize the geometric changes of the contour in local concavity / convexity, edge protrusion, or embossed areas. Subsequently, by statistically analyzing the contour height distribution, the maximum height, minimum height, average height, and height variance are extracted to form a three-dimensional height feature vector. Furthermore, by combining local slope and normal vector distribution, geometric features corresponding to the embossed texture are extracted, such as the height of the protrusion, the depth of the indentation, and the steepness of the edge.

[0037] Texture enhancement processing is performed on images of each spectral band. High-pass filtering or adaptive histogram equalization is used to enhance detailed features and suppress spectral interference from uniform background areas. Subsequently, spatial texture feature extraction methods, such as Local Binary Pattern (LBP) or Gray-Level Co-occurrence Matrix (GLCM), are combined to analyze the texture directionality, roughness, and contrast in the image, forming a pattern feature map. By fusing features from multi-band images, texture information from different bands is weighted and combined to obtain a single comprehensive pattern image that can reflect embossing, fine scratches, and wear marks. Based on the pattern outline and three-dimensional height information, local depth analysis is performed at each spatial location to calculate the height difference, curvature change, and gradient intensity of each pixel relative to the surrounding area. Through multi-scale analysis, the degree of local depressions, protrusions, and edge wear is evaluated in 3×3, 7×7, and 15×15 pixel windows. Areas with depth changes exceeding a threshold are marked as morphological defect areas. The threshold can be set within the range of 0.05 to 0.15 times the normalized height, based on the overall height variance. By combining pattern and texture information, the height differences caused by normal embossing and imprinting are eliminated, ensuring that the marking results accurately locate the real defects.

[0038] Connectivity analysis is performed on each defect area in the marked image, grouping spatially adjacent defect pixels into individual defect units. The area of ​​each defect unit is then calculated, measured in pixels, and the total number of defect units is tallied to form a defect quantity distribution. The ratio of the total defect area to the effective surface area of ​​the coin is calculated to obtain the defect area percentage. During the analysis, defect areas can be classified by size, depth, or location to generate a defect distribution density map, reflecting the distribution characteristics of defects in different materials or pattern areas. The number of defects and the defect area percentage are used as two core indicators, and a comprehensive morphological defect score is obtained through weighted fusion. The weighting can be set according to the degree of impact of defects on the condition; generally, the defect area percentage has a higher weight, set at 0.6–0.7, while the defect number has a weighting of 0.3–0.4. Normalization is performed to ensure the two indicators are within a uniform scale range, and the weighted sum is then used to form a morphological defect index.

[0039] In this embodiment, a comprehensive intelligent scoring system for currency based on multimodal data fusion is provided, used to execute the comprehensive intelligent scoring method for currency based on multimodal data fusion as described above, including: An image acquisition unit is used to acquire a set of spectral feature images of the coin to be detected; and to perform spectral texture response analysis on the set of spectral feature images to obtain a spectral texture data sequence. The defect analysis unit is used to perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map. The defect prediction unit is used to analyze material defect changes based on spectral defect response maps and generate defect change results. The condition scoring unit is used to perform a comprehensive analysis of the coin's condition based on the results of defect changes, and outputs a coin condition score.

[0040] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0041] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A comprehensive intelligent scoring method for coins based on multimodal data fusion, characterized in that, Includes the following steps: Step S1: Collect a set of spectral feature images of the coin to be detected; Spectral texture response analysis is performed on the spectral feature image set to obtain a spectral texture data sequence; Step S2: Perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map; Step S3: Analyze the material defect changes based on the spectral defect response map and generate the defect change results; Step S4: Perform a comprehensive analysis of the coin condition based on the defect changes and output the coin condition score.

2. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 1, characterized in that, The specific steps of step S1 are as follows: Collect a set of spectral feature images of the coins to be detected; Multispectral decomposition is performed on the spectral feature image set to extract multiple independent spectral signals, including material substrate components, surface texture components, and abnormal residual signals. Spectral texture response analysis is performed based on multiple independent spectral signals to obtain spectral texture data sequences.

3. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 2, characterized in that, The specific steps for acquiring the spectral feature image set of the coin to be detected are as follows: Acquire the original spectral image of the coin to be tested; Adaptive brightness equalization is performed on the original spectral image to obtain a brightness-optimized spectral image; Spectral enhancement is performed on each band of the brightness-optimized spectral image to generate a contrast-enhanced image; Spectral consistency calibration is performed on the contrast-enhanced images to obtain a set of spectral feature images.

4. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 2, characterized in that, The specific steps for performing spectral texture response analysis based on multiple independent spectral signals to obtain a spectral texture data sequence are as follows: Material substrate reflectance analysis is performed based on material substrate components to identify the spectral reflectance data of different materials; Frequency domain analysis of surface texture components yields spatial spectral distribution data; Local anomaly detection is performed on the abnormal residual signal to obtain abnormal spectral features; Dynamic spectral texture response analysis is performed on the abnormal spectral features, spatial spectral distribution data, and spectral reflectance data to obtain a spectral texture data sequence.

5. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 1, characterized in that, The specific steps of step S2 are as follows: Perform spectral defect analysis based on spectral texture data sequences and label spectral defect features; By identifying differences in response characteristics of spectral defects, multiple defect types can be obtained. Adaptive spectral channel enhancement based on multiple defect types yields spectral defect response maps.

6. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 1, characterized in that, The specific steps of step S3 are as follows: Pixel-level partitioned material recognition is performed based on the material substrate components to obtain pixel-level material recognition results; Based on the pixel-level material recognition results, region segmentation and semantic annotation are performed to obtain a material distribution semantic map; Based on the spectral defect response map, the position of the material distribution semantic map is matched, and the material defect change analysis is performed to generate the defect change results.

7. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 6, characterized in that, The defect changes include the oxidation propagation rate, corrosion depth growth rate, and crack propagation direction.

8. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 1, characterized in that, The specific steps of step S4 are as follows: The morphological defect depth visual detection is performed on the original spectral image to obtain the morphological defect index; The degree of material defect is assessed based on the results of defect changes to obtain the degree of material defect. Calculate the percentage of material degradation progress and expected remaining life based on the degree of material defects; The material wear index is generated by assessing the percentage of material degradation progress and the expected remaining lifespan. A comprehensive analysis of coin condition is conducted based on the morphological defect index and the material wear index to generate a coin condition score.

9. The comprehensive intelligent scoring method for coins based on multimodal data fusion according to claim 1, characterized in that, The specific steps for performing morphological defect depth visual detection based on the original spectral image to obtain the morphological defect index are as follows: The outer contour of the coin is identified from the original spectral image, and the three-dimensional contour line is extracted. Geometric parameters of the three-dimensional contour lines are calculated to obtain the three-dimensional morphological parameters; Surface pattern analysis was performed on the original spectral image to extract the coin pattern; Perform depth visual detection of morphological defects in coin images and 3D morphological parameters, and mark the location of morphological defects; Calculate the distribution of the number of defects and the proportion of the defect area based on the location of the morphological defects; The morphological defect index is obtained by evaluating the distribution of defect quantity and the proportion of defect area.

10. A comprehensive intelligent scoring system for coins based on multimodal data fusion, characterized in that, The method for implementing the comprehensive intelligent scoring method for coins based on multimodal data fusion as described in claim 1 includes: An image acquisition unit is used to acquire a set of spectral feature images of the coin to be detected; and to perform spectral texture response analysis on the set of spectral feature images to obtain a spectral texture data sequence. The defect analysis unit is used to perform spectral defect analysis based on the spectral texture data sequence to obtain the spectral defect response map. The defect prediction unit is used to analyze material defect changes based on spectral defect response maps and generate defect change results. The condition scoring unit is used to perform a comprehensive analysis of the coin's condition based on the results of defect changes, and outputs a coin condition score.

Images