Furniture surface paint defect detection method and system

Abstract

The present invention relates to the field of image processing and multimodal data fusion, and in particular to a method and system for detecting paint defects on furniture surfaces. The method and system perform fusion based on multimodal data, dynamically adjust the weights of different modal data through an adaptive weight allocation algorithm, and generate fused multimodal feature data. A progressive residual self-correction algorithm is used to extract multi-layer progressive defect feature data, and a feedback mechanism is used to adjust sensor acquisition parameters to perform secondary data acquisition. Finally, a multi-layer neural network is used to implement defect classification and identification, and a detection report including defect type and severity is generated. The present invention effectively improves the accuracy and efficiency of defect detection on furniture surfaces and complex areas.

Description

Technical Field

[0001] The present invention relates to the field of image processing and multimodal data fusion, and in particular to a method and system for detecting paint defects on furniture surfaces. Background Art

[0002] The purpose of painting the furniture surface is to enhance the beauty and durability of the furniture. Painting can protect the furniture from erosion by environmental factors such as light and moisture, while improving its appearance quality. However, during the painting process, due to environmental factors, improper operation or equipment defects, paint defects such as orange peel effect, sagging, powder accumulation and gap defects often appear on the furniture surface. These defects not only affect the visual effect of the furniture, but may also accelerate the aging and damage of the furniture.

[0003] The current paint defect detection mainly relies on vision-based detection methods (Chinese invention patent, publication number: CN115222743B, name: Vision-based furniture surface paint defect detection method). Although the prior art uses a paint defect detection network based on image processing, a large number of defective images and non-defective images are often required when training the network. It takes a long time to collect this data, resulting in low efficiency. In addition, these networks usually do not distinguish the importance of regions when detecting defects in various areas of the furniture surface, especially for defects in the center area of ​​the furniture, which is more susceptible to damage. The accuracy of the prior art is insufficient, making it difficult to achieve effective detection. Summary of the invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a method and system for detecting paint defects on the surface of furniture. The present invention is based on multimodal data acquisition, dynamically adjusts the weights of different modal data through an adaptive algorithm, and uses a progressive residual self-correction algorithm to extract and optimize defect feature data. The present invention can more efficiently and accurately detect paint defects on the surface and inside of furniture, especially defects in complex areas, and solves the problem of low detection accuracy of existing methods and reliance on a large amount of image data training.

[0005] A method for detecting paint defects on furniture surfaces comprises the following steps:

[0006] Acquire multimodal data of the furniture surface, wherein the multimodal data includes spectral data, sound wave reflection data and holographic imaging data, and generate preliminary multimodal fusion data;

[0007] Based on the preliminary multimodal fusion data, performing spatiotemporal characteristic fusion processing, dynamically adjusting the weights of different modal data using a multimodal adaptive weight allocation algorithm, and generating fused multimodal feature data;

[0008] Based on the fused multi-modal feature data, a progressive residual self-correction algorithm is used to perform multi-layer progressive defect feature extraction to generate verified defect feature data;

[0009] Based on the verified defect feature data, adaptive feedback and dynamic adjustment are performed, sensor parameters for collecting multimodal data are adjusted through feedback residual analysis, and a second data collection is performed to generate second multimodal fusion data;

[0010] Based on the second multimodal fusion data, a multi-layer neural network is used to perform defect recognition and classification, to generate defect classification data including defect type and severity, and to generate a paint defect detection report based on the defect classification data.

[0011] Preferably, the preliminary processing of the multimodal data includes noise filtering of spectral data, acoustic reflection data and holographic imaging data, and the filtering step uses a preset noise threshold value to remove environmental interference and sensor errors; the data of different modalities are standardized through a normalization processing algorithm, so that the data of each modality are compared and fused within a unified numerical range, thereby generating preliminary multimodal fusion data.

[0012] Preferably, the multimodal adaptive weight allocation algorithm dynamically adjusts the weights of different modal data by the following formula:

[0013]

[0014] in, For the The weight of the modal data; For the The accuracy of the modal data; is the total number of modal data.

[0015] Preferably, the accuracy Calculated by the following formula:

[0016]

[0017] in, For the The accuracy of the modal data; For the The number of sampling points of each modal data; For the The modal data is in The value of the data point; For the The average value of all modal data points.

[0018] Preferably, the progressive residual self-correction algorithm corrects the residual in defect feature extraction by the following formula:

[0019]

[0020] in, is the current corrected residual value; is the residual value of the previous round; is the preset learning rate; The value extracted for the current feature.

[0021] Preferably, the progressive residual self-correction algorithm optimizes the accuracy of defect feature extraction through multiple rounds of correction, including the following steps: in each round of residual correction, the extracted defect features are analyzed, and the residuals are fed back to the feature extraction model, the model parameters are adjusted, and finally optimized defect feature data is generated, and the defect feature data is used to identify the defect location and severity on the surface and interior of the furniture.

[0022] Preferably, in the step of adaptive feedback and dynamic adjustment, the acquisition parameters of the multi-frequency spectrum sensor, the acoustic wave sensor and the holographic imaging sensor are dynamically adjusted through feedback residual analysis, and the specific adjustment includes:

[0023] Adjust the wavelength range of the multi-frequency spectral sensor to capture surface color changes;

[0024] Adjust the amplitude and frequency of the acoustic sensor to detect defects inside furniture;

[0025] Adjust the resolution of holographic imaging sensors to improve the detection accuracy of surface texture and structure.

[0026] Preferably, the second data acquisition is performed based on dynamically adjusted sensor parameters to generate second multimodal fusion data, and the second multimodal fusion data is used to improve the accuracy of defect detection on complex surfaces and internal parts of furniture.

[0027] Preferably, the multi-layer neural network adjusts the parameters and weights of the model through a feedback enhancement mechanism, and the feedback enhancement mechanism includes the following steps: in each round of defect identification, the identified defect classification results are fed back to the neural network, and the weights of the neural network are adjusted through the feedback to optimize the classification accuracy, and finally generate defect classification data and repair suggestions.

[0028] A system for implementing the furniture surface paint defect detection method, the system comprising:

[0029] A multimodal data acquisition module, used to acquire multimodal data of the furniture surface, wherein the multimodal data includes spectral data, acoustic wave reflection data and holographic imaging data acquired by a multi-frequency spectral sensor, an acoustic wave sensor and a holographic imaging sensor;

[0030] A data processing module, used to perform preliminary processing on the multimodal data, including noise filtering and normalization processing, and generate preliminary multimodal fusion data;

[0031] A spatiotemporal characteristic fusion module, for performing spatiotemporal characteristic fusion processing based on the preliminary multimodal fusion data, dynamically adjusting the weights of different modal data using a multimodal adaptive weight allocation algorithm, and generating fused multimodal feature data;

[0032] A defect feature extraction module, used to perform multi-layer progressive defect feature extraction based on the fused multi-modal feature data using a progressive residual self-correction algorithm to generate verified defect feature data;

[0033] A feedback and dynamic adjustment module, which is used to perform adaptive feedback and dynamic adjustment based on the verified defect feature data, adjust the parameters of the multi-frequency spectrum sensor, the acoustic wave sensor and the holographic imaging sensor through feedback residual analysis, and perform a second data acquisition to generate second multimodal fusion data;

[0034] A defect recognition and classification module is used to perform defect recognition and classification based on the second multimodal fusion data using a multi-layer neural network, generate defect classification data including defect type and severity, and generate a paint defect detection report based on the defect classification data.

[0035] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0036] The present invention expands the depth and breadth of defect identification, realizes efficient data fusion, and improves the accuracy of detection through multimodal data acquisition, that is, acquisition of multiple data such as spectrum, acoustic wave reflection and holographic imaging;

[0037] The present invention adopts an adaptive weight allocation algorithm to achieve dynamic adjustment of weights of different data modes, overcome the limitations of single visual detection, enable different modal information to complement each other in defect detection, and improve detection accuracy;

[0038] The present invention uses a progressive residual self-correction algorithm, and the present invention can perform feedback adjustment in each round of defect feature extraction process to further optimize the detection accuracy, especially for defect detection in complex areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the process of the present invention;

[0040] Figure 2 It is a schematic diagram of multimodal data collection in the present invention;

[0041] Figure 3 A schematic diagram of weight adjustment of the adaptive weight allocation algorithm in the present invention;

[0042] Figure 4 It is a working flow chart of the progressive residual self-correction algorithm in the present invention;

[0043] Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0044] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0045] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0046] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0047] like Figure 1 As shown, a method for detecting paint defects on furniture surface comprises the following steps:

[0048] Acquire multimodal data of the furniture surface, wherein the multimodal data includes spectral data, sound wave reflection data and holographic imaging data, and generate preliminary multimodal fusion data;

[0049] like Figure 2As shown, preferably, the preliminary processing of the multimodal data includes noise filtering of spectral data, acoustic reflection data and holographic imaging data, and the filtering step uses a preset noise threshold value to remove environmental interference and sensor errors; the data of different modalities are standardized through a normalization processing algorithm, so that each modal data is compared and fused within a unified numerical range, thereby generating preliminary multimodal fusion data.

[0050] Multimodal data includes spectral data, acoustic reflection data, and holographic imaging data, which are collected through a variety of sensors and correspond to different physical properties of the furniture surface. Spectral data is used to detect color differences and thickness changes in the furniture surface coating, acoustic reflection data is used to detect structural defects (such as bubbles or cracks) under the furniture surface, and holographic imaging data captures the three-dimensional morphological information of the furniture surface, such as surface unevenness or other minor geometric deformations. By acquiring these three types of data at the same time, the quality of furniture surface paint can be comprehensively analyzed from multiple angles, avoiding the limitations of a single detection method.

[0051] After data acquisition, the system will perform preliminary processing of multimodal data. This processing includes noise filtering and normalization. First, the purpose of noise filtering is to remove external interference, such as lighting changes, environmental vibrations, or sensor errors, through a preset noise threshold. Here, the noise threshold can be set by experimental data or industry standards to ensure that only valid detection data is retained and reduce the interference of errors on the detection results. The principle of this step is based on the signal-to-noise ratio principle in signal processing, and the accuracy and stability of the overall detection are improved by filtering out data with low signal-to-noise ratio. For example, when detecting the surface of furniture, if there are background light changes from the environment or thermal noise from the device itself, these interference signals will affect the accuracy of spectral reflection data. Through noise filtering, these external influences can be effectively reduced, so that the spectral data finally detected can more truly reflect the actual state of the furniture surface coating. Similarly, for acoustic reflection data, noise filtering can remove invalid data caused by environmental vibration, so that the system can more accurately detect subtle changes in the internal structure of furniture.

[0052] Next is normalization, which is to standardize the data of different modalities so that they can be compared and fused within a unified numerical range. Since the dimensions and numerical ranges of spectral, acoustic and holographic imaging data are different, if these data are directly processed and fused, the influence of a certain modality on the final result will be too large or too small, resulting in distortion or deviation. Normalization can linearly transform or nonlinearly map data of different modalities to ensure that all types of data have the same basis for comparison on a unified numerical scale, thereby providing high-quality data input for subsequent multimodal fusion. For example, the effect of normalization is that even if the numerical range of spectral data is small (such as 0-1), and holographic imaging data may be in the millimeter range, after normalization, both are mapped to the same range, thereby avoiding data imbalance caused by numerical differences during data fusion.

[0053] Through the above steps, preliminary multimodal fusion data is generated, which provides a unified and optimized input data set for subsequent spatiotemporal feature fusion and defect feature extraction. Such a data processing flow not only ensures the consistency of each modal data, but also significantly improves the accuracy and robustness of defect detection. In practical applications, this multimodal data fusion method can effectively identify subtle paint defects on the surface of furniture and potential internal structural problems. For example, by comparing the test results before and after processing, it can be found that multimodal data after noise filtering and normalization is more likely to detect subtle color differences, uneven coating thickness, and difficult-to-detect internal bubbles.

[0054] Based on the preliminary multimodal fusion data, performing spatiotemporal characteristic fusion processing, dynamically adjusting the weights of different modal data using a multimodal adaptive weight allocation algorithm, and generating fused multimodal feature data;

[0055] In the present invention, the spatiotemporal characteristic fusion processing is performed based on the preliminary multimodal fusion data in order to more accurately extract the paint defect information on the furniture surface. The core principle of the spatiotemporal characteristic fusion processing is to integrate the spatial and temporal features of different modalities (spectrum, acoustic reflection and holographic imaging) so that the multimodal data can better reflect the comprehensive information of the furniture surface and its internal structure. Since each modal data has different sensitivity and response characteristics, for example, spectral data is more sensitive to changes in the color and thickness of the surface coating, while acoustic reflection data is good at detecting hidden defects under the coating, the key to spatiotemporal characteristic fusion is how to synchronously integrate these different modal data in the time and space dimensions to obtain a more comprehensive description of surface defects.

[0056] When performing spatiotemporal feature fusion, the present invention adopts a multimodal adaptive weight allocation algorithm, which is used to dynamically adjust the weight of each modality according to the accuracy, correlation and noise level of different modal data. The core principle of the adaptive weight allocation algorithm is that for each detection point or detection area, the system will analyze the validity of each modality data in real time, and automatically assign appropriate weights to each modality to ensure that in the final fusion, the higher quality data has a greater impact on the result, while the data with greater noise or lower precision has less impact. Specifically, the system will calculate the accuracy and noise level of each modality data, and the data with higher accuracy and lower noise will obtain higher weights. For example, on some highly reflective furniture surfaces, holographic imaging data may be noisy due to changes in reflection intensity caused by surface gloss. At this time, the weight allocation algorithm will reduce the weight of the holographic data and assign more weight to the spectral and acoustic reflection data because these data are more reliable in this scenario.

[0057] Through this dynamic adjustment mechanism, the spatiotemporal characteristics of each modal data can be fused to generate fused multimodal feature data. This feature data not only combines the surface color, thickness and structure information, but also integrates the changes of each modality in different spatiotemporal dimensions, making subsequent paint defect detection more accurate. For example, there is a tiny crack somewhere on the surface of the furniture that is difficult to detect with the naked eye. The spectral data may show that the coating color in this area has changed slightly, while the sound wave data shows that the amplitude of this area is abnormal. Through the weight distribution algorithm, these two types of information are effectively integrated, allowing the system to accurately locate the area and generate accurate defect feature data.

[0058] The effect of this adaptive weight distribution is reflected in the ability to handle complex detection scenarios. In traditional single-modality detection, noise or failure in a certain mode may lead to inaccurate detection, but through the fusion of spatiotemporal characteristics, the system can intelligently select the optimal modality data, avoiding the limitations of single-modality detection. For example, in actual operation, when detecting furniture surfaces with different gloss or complex textures, the holographic imaging of certain areas may be disturbed by reflected light, while in the same area, the acoustic reflection data still has a high degree of accuracy, so the weight of the holographic data in this area will automatically decrease, and the weight of the acoustic data will increase, thereby ensuring the robustness and reliability of defect detection.

[0059] like Figure 3 As shown, preferably, the multimodal adaptive weight allocation algorithm dynamically adjusts the weights of different modal data through the following formula:

[0060]

[0061] in, For the The weight of the modal data; For the The accuracy of the modal data; is the total number of modal data.

[0062] The core purpose of the multimodal adaptive weight allocation algorithm is to dynamically adjust the weight of each modal data in the overall fusion according to the accuracy of different modal data. For the detection of paint defects on furniture surfaces, each modal data (such as spectral data, acoustic reflection data, and holographic imaging data) has different characteristics and application scenarios. Therefore, it is necessary to assign appropriate weights to each modal data according to the specific detection situation to ensure the accuracy and reliability of the final fused data. By dynamically adjusting the weights, the detection accuracy can be improved, especially when dealing with complex surfaces, different textures, and thickness changes. The algorithm can automatically optimize the data fusion process according to environmental changes.

[0063] Indicates The weight of the modal data, It is The accuracy of the modal data is is the total number of modal data. The purpose of this formula is to ensure that when the sum is 1, each modal data is assigned a corresponding weight according to its relative accuracy. Here, accuracy Represents the quality of the modality data, usually calculated based on its signal-to-noise ratio or other indicators. For high-precision data, the algorithm assigns it a higher weight, making this data have a greater impact on the final fusion result; for low-precision data, the algorithm will reduce its weight accordingly, reducing its impact on the overall fusion result.

[0064] The effect is reflected in coping with the complexity and diversity of the actual detection environment. For example, when detecting the surface of furniture, spectral data can usually provide information about the color and thickness of the coating, but the accuracy of the spectral data may decrease when the light is complex or the surface reflection is too strong. Holographic imaging data has higher accuracy in capturing surface morphology, so in this case, the algorithm will automatically increase the weight of holographic imaging data and reduce the weight of spectral data. This dynamic adjustment enables the system to always maintain a high detection accuracy under various complex conditions, avoiding the serious impact of the failure of a certain modal data on the overall detection results.

[0065] Specifically, suppose that when detecting an area on the surface of furniture, the spectral data is affected by the ambient light, resulting in reduced detection accuracy. The system can identify this through an adaptive weight allocation algorithm and allocate more weight to the sound wave reflection data and holographic imaging data, which have higher data accuracy in this scenario. Therefore, although the spectral data is affected, by dynamically adjusting the weights, the system can still accurately detect the paint defects in the area, avoiding errors caused by the failure of a certain data source.

[0066] Through this algorithm, the effect has been significantly improved, especially when facing furniture surfaces with high gloss, complex geometry or uneven coating. The algorithm can automatically optimize the weight distribution according to the data quality to ensure that the final generated multimodal fusion data is more reliable and accurate. In addition, this dynamic adjustment mechanism also improves the robustness of the system, enabling it to adapt to a variety of operating environments and detection conditions, and still maintain high detection performance when dealing with surface reflections, sensor errors or environmental interference.

[0067] Preferably, the accuracy Calculated by the following formula:

[0068]

[0069] in, For the The accuracy of the modal data; For the The number of sampling points of each modal data; For the The modal data is in The value of the data point; For the The core idea of ​​this formula is to measure the stability and accuracy of the modal data by calculating the deviation between the modal data of each data point and the average value of all modal data.

[0070] Accuracy of multimodal data It is a key indicator used to evaluate the accuracy and reliability of each modal data at each detection point. The accuracy directly affects the weight distribution of each modal data in the subsequent fusion step, thus determining the accuracy and robustness of the final detection result.

[0071] In the application of the present invention, the spectral data, the acoustic reflection data and the holographic imaging data respectively reflect the different physical properties of the furniture surface. Through the above formula, the system can calculate the accuracy of these modal data at each detection point. The basic principle of the formula is: if the difference between a certain modal data and other modal data is small, it means that the consistency of the modal data with the overall data is high, and its accuracy is On the contrary, if a modal data deviates greatly from the average value, its accuracy is low. In other words, the formula is based on the absolute value of the difference. To measure the deviation of the data, the deviation is normalized to ensure that the deviations between different data points can be compared on the same dimension.

[0072] For example, suppose the system acquires three types of data at a certain detection point: spectrum, sound wave, and holographic imaging. Average with other modes A large difference means that the spectral data may be affected by ambient light or equipment problems, resulting in a large error. In this case, the formula will calculate the difference and obtain a lower spectral data accuracy. , thereby reducing the weight of the modality data in the subsequent multimodal fusion to prevent its erroneous data from excessively affecting the final detection results. and The difference between them is small, indicating that the acoustic wave data is more stable and reliable at this detection point, and the formula will assign a higher accuracy to the acoustic wave data. , making it have a higher weight in the fusion process.

[0073] The significant effect is reflected in the improvement of the system's adaptability and robustness. In the actual detection process, the furniture surface often leads to differences in data quality of different modes due to its material, coating thickness and the complexity of the external environment. Through the precision formula, the system can automatically identify which modal data performs better in a specific environment and dynamically adjust the weight of this data, thereby maximizing the use of high-precision data and reducing the negative impact of low-precision data on the detection results. For example, when the system detects that the spectral data of a certain furniture surface fluctuates due to its high gloss, the formula automatically reduces the accuracy of the spectral data and increases the influence of sound waves or holographic data. In this way, even if there is noise in the spectral data, the system can still ensure the detection accuracy through other more reliable data sources.

[0074] Through this mechanism, the system can effectively cope with various complex furniture surface paint defect detection scenarios. Whether it is a high-gloss surface, complex texture structure or uneven coating, the system can adaptively adjust according to the accuracy of each modal data, thereby ensuring the accuracy and stability of the detection results.

[0075] like Figure 4 As shown, based on the fused multi-modal feature data, a progressive residual self-correction algorithm is used to perform multi-layer progressive defect feature extraction to generate verified defect feature data;

[0076] Preferably, the progressive residual self-correction algorithm corrects the residual in defect feature extraction by the following formula:

[0077]

[0078] in, is the current corrected residual value; is the residual value of the previous round; is the preset learning rate; The value extracted for the current feature.

[0079] This formula reflects that in each round of iteration, by The gradient (partial derivative) ) is calculated and multiplied by a learning rate , correct the current residual value , thereby gradually reducing the residual and making the error of feature extraction continuously reduced.

[0080] In the actual application of furniture surface paint inspection, different defect features (such as uneven coating thickness, bubbles, cracks, etc.) are captured and extracted through multimodal data. These feature data may cause deviations due to data noise or sensor accuracy problems when they are extracted for the first time. At this time, the system calculates the residual of this feature extraction, that is, the difference between the extracted features and the expected features. Through the progressive residual self-correction algorithm, the system uses the gradient information of the previous round of residuals for correction and adjusts the current feature extraction value, so that the system's capture of defect features gradually approaches the true value. For example, when detecting tiny cracks on the surface of furniture, the initial extraction may cause data distortion due to surface reflection. By correcting the residuals through multiple iterations, the system can more accurately describe the actual size and position of the cracks.

[0081] The progressive residual self-correction algorithm greatly improves the accuracy and robustness of the detection results, especially in complex environments or when there are differences in multimodal data, its optimization effect is particularly obvious. Through the progressive correction of residuals, the system can overcome the uncertainty caused by noise and errors in the data acquisition process, thereby ensuring that the final extracted defect features are more accurate and reliable. In traditional defect detection methods, once data deviation occurs, the system is often difficult to correct, resulting in large errors in the final detection results. Through progressive residual self-correction, the system can correct the error in each round of iteration, and finally converge to a smaller residual, making the defect features closer to the true value.

[0082] For example, suppose there is a tiny dent on the surface of the furniture detected. The dent may be misjudged by the system as a small area of ​​uneven coating during the initial extraction. With the iteration of the progressive residual self-correction algorithm, the system continuously adjusts the feature extraction results based on the gradient information of the residual after each calculation of the residual, and gradually corrects the defect description of the area. After multiple rounds of correction, the system can more accurately determine the true shape of the dent, including information such as depth, area and location, thereby providing more accurate data support for subsequent paint repairs.

[0083] The progressive residual self-correction algorithm optimizes the errors in each round of feature extraction in a progressive manner, enabling the system to obtain more accurate and reliable defect features in multimodal data. This algorithm not only improves the accuracy of the detection results, but also enhances the system's adaptability and stability to complex detection environments, ensuring that furniture surface paint defect detection can obtain consistent and reliable results under different conditions.

[0084] Preferably, the progressive residual self-correction algorithm optimizes the accuracy of defect feature extraction through multiple rounds of correction, including the following steps: in each round of residual correction, the extracted defect features are analyzed, and the residuals are fed back to the feature extraction model, the model parameters are adjusted, and finally optimized defect feature data is generated, and the defect feature data is used to identify the defect location and severity on the surface and interior of the furniture.

[0085] In the present invention, the main purpose of the progressive residual self-correction algorithm is to optimize the accuracy of defect feature extraction through multiple rounds of iterations to ensure that the defect location and severity of the furniture surface and interior are accurately captured in multimodal data. The algorithm analyzes the defect features extracted in each round, feeds the residuals back to the feature extraction model, and continuously corrects the extraction results by adjusting the model parameters, so that the defect feature data finally generated is more accurate.

[0086] The core idea of ​​the progressive residual self-correction algorithm is residual feedback and iterative optimization. After each round of defect feature extraction, the system calculates the gap between the feature data and the expected target to generate a residual value. The residual represents the deviation between the current extraction result and the actual defect. A larger residual means that there is a larger error in the extraction result. In order to reduce this error, the system will automatically adjust the model parameters based on the residual information feedback to the feature extraction model. This adjustment process involves optimizing key parameters such as the model weight and learning rate to ensure that the model can extract defect features more accurately in subsequent rounds. By continuously correcting the model, the system can gradually reduce the residual so that the extraction result converges to a value close to the actual defect feature after multiple rounds of correction.

[0087] For example, suppose that when detecting the thickness of the paint on the surface of furniture, the initially extracted feature data may cause the system to misjudge the area with uneven paint thickness due to the influence of sensor accuracy or environmental noise. After this round of extraction, the system calculates the residual between the current extraction result and the actual expected value, and feeds the residual back to the model. After receiving the residual information, the model adjusts its parameters, such as reducing the learning rate or increasing the weight of specific data, to better fit the real features. In the next iteration, the system will extract the paint thickness information of the area again, and through further analysis and correction, gradually narrow the gap with the real value.

[0088] The effect of this multiple-round correction is that it significantly improves the accuracy and stability of defect detection, especially when detecting complex surfaces. For example, when there are complex textures or curved areas on the surface of furniture, a single data extraction may not be able to accurately capture the paint defects on these surfaces, which may lead to oversimplification of feature extraction results or misjudgment of certain tiny depressions and protrusions as serious defects. The progressive residual self-correction algorithm uses multiple rounds of residual feedback to enable the system to more accurately distinguish these complex structures and gradually extract defects such as uneven coating thickness, cracks or bubbles.

[0089] In practice, the algorithm gradually optimizes the feature extraction results of each round through multiple iterations. Suppose that in the process of detecting small bubbles on the surface of a piece of furniture, the features extracted for the first time may be inaccurate in judging the position and size of the bubbles due to excessive noise or insufficient data. Under the residual self-correction mechanism, the system feeds back the data deviation of the bubble area to the model, adjusts the data weight and acquisition parameters of the specific area, and re-extracts and analyzes the area in the next iteration. Through multiple corrections, the system can finally accurately determine the true position and size of the bubble, and provide accurate data information for subsequent repairs.

[0090] The advantage of the application is that the progressive residual self-correction algorithm can adapt to various complex detection scenarios and automatically optimize the extraction results, avoiding misjudgment caused by the accumulation of errors in one-time detection. When detecting minor defects such as coating thickness changes, internal cracks, and surface depressions, the system uses this algorithm to ensure more accurate capture of details, effectively improving the overall reliability and accuracy of paint defect detection.

[0091] The progressive residual self-correction algorithm achieves the gradual optimization of furniture surface paint defect feature extraction through residual feedback and multiple rounds of correction. It is particularly effective in the case of complex surfaces and large differences in modal data. Through multiple rounds of parameter adjustment, it ensures that the system can accurately identify the defect location, type and severity, providing accurate reference data for subsequent repair work.

[0092] Based on the verified defect feature data, adaptive feedback and dynamic adjustment are performed, sensor parameters for collecting multimodal data are adjusted through feedback residual analysis, and a second data collection is performed to generate second multimodal fusion data;

[0093] In the present invention, adaptive feedback and dynamic adjustment are a key optimization mechanism, which further improves the accuracy of multimodal data acquisition based on verified defect feature data. During the detection process, the multimodal data initially collected by the system is analyzed and processed to generate preliminary defect feature data, but due to the complexity of the data or the noise problem of the sensor, the data collected initially may have some inaccurate areas, especially in some cases where the coating thickness is uneven, the surface texture is complex, or the internal defects of the furniture are relatively hidden. Therefore, it is necessary to dynamically adjust the parameters of the sensor through adaptive feedback in order to obtain more accurate multimodal data in the second data collection, and finally generate more reliable defect feature information.

[0094] Adaptive feedback means that after the initial feature extraction is completed and the verified defect feature data is generated, the system will analyze the residuals in these data, that is, the difference between the extracted features and the expected results. Through this residual analysis, the system can identify problem areas in the data, such as the low accuracy of some sensor data in a specific area, or the inability of a certain modal data to effectively reflect the real defect due to external interference. Based on this residual information, the system automatically adjusts the sensor parameters used to collect multimodal data, including adjusting the wavelength range of the spectral sensor, the frequency and amplitude of the acoustic sensor, and the resolution of the holographic imaging sensor. The purpose of these parameter adjustments is to improve the sensitivity and accuracy of the sensor under specific conditions, so as to more accurately capture defect features.

[0095] Suppose that when detecting a certain furniture surface, the spectral data is deviated from the initial data collection due to the strong reflection of light on the surface, and the acoustic wave data is affected by the surface roughness, and the signal reflection is unstable, resulting in the system being unable to accurately identify the uneven thickness of the paint in a certain area. Through the adaptive feedback mechanism, the system will identify the problem areas of these data and automatically adjust the operating parameters of the sensor through feedback residual analysis. For example, the wavelength range of the spectral sensor can be adjusted to a band that is more sensitive to changes in surface color and thickness to reduce the interference of reflected light on the data; at the same time, the acoustic wave sensor can better adapt to the structural changes of complex surfaces by adjusting the amplitude and frequency. These dynamic adjustments enable the system to obtain multimodal data with higher accuracy in the second data collection.

[0096] The purpose of the second data collection is that, by optimizing and adjusting the sensor parameters, the system can re-collect data in areas with lower accuracy before, ensuring that the data quality in these areas is improved, thereby generating more accurate multimodal fusion data. These new data will contain defect features that were not accurately captured before, or improve the accuracy of identified defect features. For example, after the second data collection, the system may re-identify a certain area where the coating thickness changes, and due to the adjustment of the wavelength range of the spectral sensor, this data can more accurately reflect the thickness changes of the coating, thereby more accurately identifying the size and location of the defect.

[0097] Through this dynamic adjustment mechanism, the system can not only make up for the deficiencies in the initial data collection, but also further optimize the defect feature data that has been generated. The advantage of this mechanism is that it does not rely on the results of a single collection, but gradually improves the system's detection capabilities through a cyclic feedback and adjustment process. For furniture surfaces with complex structures, especially highly reflective surfaces, areas with complex textures, or deep hidden defects, a single inspection is often difficult to capture all the details. Through the adaptive feedback mechanism, the system can identify these difficult-to-detect areas and optimize the results in multiple collections by dynamically adjusting sensor parameters, ultimately generating more accurate and comprehensive defect feature data.

[0098] In addition, the adaptive feedback mechanism can also improve the robustness of the system and adapt to various operating environments and detection conditions. For example, in an environment with uneven light, the system can reduce the impact of light changes on the detection results by adjusting the spectral sensor parameters; when facing furniture surfaces with large variations in coating thickness, the dynamic adjustment of the acoustic sensor can help more accurately identify thickness changes and avoid misjudgment. Through these multiple rounds of optimization and feedback, the system can not only ensure high accuracy of detection results under different conditions, but also effectively improve the ability to identify various types of defects.

[0099] Adaptive feedback and dynamic adjustment optimize the system's detection capabilities in multiple rounds of data collection by dynamically adjusting the sensor's parameters. Through feedback residual analysis, the system can identify problem areas in the detection and adjust the acquisition parameters in a targeted manner to ensure more accurate multimodal data in the second data collection. Ultimately, this process significantly improves the detection accuracy of paint defects on furniture surfaces, especially in the detection of defects on complex surfaces or structures that are more hidden, and can produce more accurate and reliable detection results.

[0100] Preferably, in the step of adaptive feedback and dynamic adjustment, the acquisition parameters of the multi-frequency spectrum sensor, the acoustic wave sensor and the holographic imaging sensor are dynamically adjusted through feedback residual analysis, and the specific adjustment includes:

[0101] Adjust the wavelength range of the multi-frequency spectral sensor to capture surface color changes;

[0102] Adjust the amplitude and frequency of the acoustic sensor to detect defects inside furniture;

[0103] Adjust the resolution of holographic imaging sensors to improve the detection accuracy of surface texture and structure.

[0104] In the present invention, the adaptive feedback and dynamic adjustment step improves the accuracy and reliability of data acquisition by dynamically adjusting the parameters of multimodal sensors (including multi-frequency spectral sensors, acoustic wave sensors, and holographic imaging sensors). The core of this process is to analyze the feedback residuals based on the previous detection results, identify the situation where the sensor parameters are insufficient to capture certain details or specific area features, and adjust the working parameters of the relevant sensors to further optimize the defect detection effect.

[0105] Feedback residual analysis generates residual data by comparing the expected value of sensor data with the actual detection value. Through these residual data, it is possible to identify which sensor parameters perform poorly under certain detection conditions, so that the system can optimize the sensor parameter settings in real time. Specifically, in the present invention, the adjustment of sensors focuses on the following three aspects: wavelength range adjustment of multi-frequency spectral sensors, amplitude and frequency adjustment of acoustic wave sensors, and resolution adjustment of holographic imaging sensors.

[0106] Wavelength range adjustment of multi-frequency spectral sensor: Spectral data is mainly used to capture color changes and paint thickness differences on furniture surfaces. Surfaces of different colors and materials have different reflectivity responses to spectral wavelengths, so adjusting the wavelength range of the sensor can better adapt to the characteristics of specific furniture surfaces. After the initial data collection, if the feedback residual analysis shows that color changes in certain areas are not fully reflected (such as subtle color differences on high-gloss surfaces), the system will automatically adjust the wavelength range of the spectral sensor to a band that is more sensitive to color changes, thereby improving the detection accuracy of that area. For example, for dark wood surfaces, the short wavelength spectrum may not be able to effectively capture color changes. At this time, the system can adjust the sensor wavelength to the long wave range to enhance the detection capability of dark surfaces.

[0107] Amplitude and frequency adjustment of acoustic wave sensors: Acoustic wave reflection data is mainly used to detect defects inside furniture, such as bubbles, cracks, or changes in coating thickness. Acoustic wave sensors send and receive sound waves to analyze their reflection and propagation on the surface or inside of furniture. Different materials and defect types have different reflection and absorption characteristics for sound waves. By adjusting the amplitude and frequency of the acoustic wave sensor, internal defects can be better detected. For thicker coatings or furniture with multi-layer structures, higher frequency sound waves may not be able to penetrate deep layers, while lower frequency sound waves can better detect hidden defects inside. Therefore, when the feedback residual analysis finds abnormal sound wave reflection or insufficient detection depth in certain areas, the system adjusts the frequency and amplitude of the sensor to adapt to the detection needs of that area. For example, when detecting thick wooden furniture, the system can detect internal cracks or gaps by lowering the frequency of the sound waves.

[0108] Resolution adjustment of holographic imaging sensors: Holographic imaging data is used to capture the three-dimensional texture structure of furniture surfaces, especially when detecting complex surface morphology (such as uneven surfaces or subtle paint defects). Holographic imaging has great advantages. The resolution of the holographic imaging sensor directly determines the system's capture accuracy of the surface microstructure. During the inspection process, if the system finds through feedback residual analysis that certain subtle defects (such as small areas of uneven coating or tiny surface scratches) are not accurately extracted, the system will dynamically adjust the resolution of the holographic sensor to improve the ability to capture image details. For example, when dealing with high-gloss paint surfaces, holographic imaging sensors may require higher resolution to capture subtle reflective changes and uneven areas on the surface.

[0109] Through the dynamic adjustment of the above three types of sensor parameters, the system can generate a second multimodal fusion data, which is more accurate than the data collected initially, especially in areas with large errors or missing data. The new sensor parameter settings enable the system to re-collect high-quality data. For example, if the spectral data collected by the system for the first time leads to inaccurate color judgment due to light reflection, or the sound wave reflection data fails to effectively detect deep defects, after parameter adjustment, the system can correct these problems in the second data collection, thereby generating more accurate defect feature data. This not only improves the accuracy of the entire detection process, but also provides a more reliable basis for subsequent defect repair and quality control.

[0110] Through adaptive feedback and dynamic adjustment, the system has greatly enhanced its adaptability to complex detection scenarios. Different furniture surface materials, coating thicknesses, and structural characteristics have different requirements for sensors. Traditional static parameter settings often cannot cover all detection needs. The present invention uses feedback residual analysis and dynamic adjustment to not only meet the surface detection needs of different materials, but also improve the detection depth and accuracy of hidden defects inside furniture. Ultimately, the dynamically adjusted multimodal data acquisition significantly improves the overall detection accuracy of the system, especially when dealing with high-gloss, complex textures, or multi-layered furniture.

[0111] Preferably, the second data acquisition is performed based on dynamically adjusted sensor parameters to generate second multimodal fusion data, and the second multimodal fusion data is used to improve the accuracy of defect detection on complex surfaces and internal furniture.

[0112] The second data collection is based on dynamically adjusted sensor parameters, aiming to generate more accurate multimodal fusion data through a more optimized collection method, thereby improving the accuracy of detecting defects on complex surfaces and inside furniture. After the first data collection and feedback residual analysis, the system can identify which sensor parameters fail to fully capture defect characteristics. Therefore, in the second data collection, the system will adjust the key parameters of sensors such as spectrum, sound wave and holographic imaging to correct the deviations or blind spots in the initial collection.

[0113] The core of the second data collection is to use the results of the first data collection for feedback optimization and dynamically adjust the sensor's collection conditions. For example, the wavelength range of the spectral sensor, the amplitude and frequency of the acoustic sensor, the resolution of the holographic imaging sensor, etc. may all be adjusted based on the analysis results of the feedback residual. This adjustment is to better adapt to the material, morphology, and complexity of the coating on the surface of a specific piece of furniture. For example, if the data collected for the first time shows that the spectral sensor is not sensitive enough to capture color changes on certain high-gloss surfaces, the system may increase the spectral wavelength range in that area; and for the acoustic sensor, if the detected data fails to deeply detect the internal structure of the furniture, the system will reduce the frequency of the acoustic wave in order to detect deep internal defects.

[0114] When inspecting a piece of furniture with a complex structure, the first data collection may result in insufficient resolution of the holographic imaging data due to reflections on the furniture surface or changes in paint thickness, and it may not be able to accurately capture subtle changes in surface texture. After feedback residual analysis, the system will automatically adjust the resolution of the holographic imaging sensor so that it can more clearly capture surface concave and convex changes in the second data collection, especially in areas with complex textures. In addition, the frequency adjustment of the acoustic wave sensor may also be used to detect deeper defects inside the furniture, such as tiny cracks or bubbles, especially in areas that were not fully detected in the first data collection. Through the adjusted second data collection, the system can obtain higher resolution and deeper information.

[0115] The second data collection after dynamic adjustment significantly improves the accuracy of detecting defects on complex surfaces and inside furniture. During the first data collection process, the preset parameters of various sensors may not be fully adapted to the specific detection scenario, resulting in the failure to fully detect defects in some areas; through the second collection, the system can optimize the sensor parameters in a targeted manner and re-collect these problem areas to obtain more accurate multimodal fusion data. These optimized data not only improve the detection accuracy of surface paint defects, but also enable the system to better capture hidden defects inside furniture, especially complex surfaces, uneven coating thickness, or complex internal structures.

[0116] For example, there may be tiny cracks or changes in paint thickness in a local area on the surface of furniture, and the spectral data collected the first time failed to fully reflect these changes due to light reflection or surface roughness. After feedback residual analysis, the system adjusted the wavelength range of the spectral sensor and collected data again. The new spectral data more accurately captured the subtle color differences and thickness changes in the area, thereby helping the system to better identify potential defects. Similarly, the acoustic wave sensor may not detect deeper cracks inside the furniture during the initial collection due to the high frequency setting. By lowering the acoustic wave frequency during the second collection, the system was able to detect deeper internal problems and generate more detailed and reliable defect feature data.

[0117] The second multimodal fusion data combines the optimized data of spectrum, sound wave and holographic imaging. Through dynamically adjusted multimodal fusion, the system can more accurately detect complex surface and internal defects, ensuring that the defect feature data finally generated has high resolution and reliability. This data can not only be used for further analysis and evaluation, but also provide a reliable basis for subsequent repair suggestions. For example, on the surface of a piece of furniture with complex carvings, the adjusted holographic imaging data can more accurately capture the subtle depressions and uneven paint spraying in the carved area, making the defect detection report generated by the system more detailed and providing clear guidance for the repair work.

[0118] The second data collection generates more accurate multimodal fusion data by dynamically adjusting the sensor parameters. This data collection step not only significantly improves the detection accuracy of paint defects on the furniture surface, but also enhances the system's ability to identify complex furniture surface and internal defects, making the entire detection process more comprehensive and reliable, and providing high-quality data support for the final detection report and repair plan.

[0119] Based on the second multimodal fusion data, a multi-layer neural network is used to perform defect recognition and classification, to generate defect classification data including defect type and severity, and to generate a paint defect detection report based on the defect classification data.

[0120] Preferably, the multi-layer neural network adjusts the parameters and weights of the model through a feedback enhancement mechanism, and the feedback enhancement mechanism includes the following steps: in each round of defect identification, the identified defect classification results are fed back to the neural network, and the weights of the neural network are adjusted through the feedback to optimize the classification accuracy, and finally generate defect classification data and repair suggestions.

[0121] In the present invention, based on the second multimodal fusion data, the use of a multi-layer neural network for defect recognition and classification is one of the key links in the entire detection process. After multimodal data fusion, the system can obtain comprehensive data on the surface and interior of the furniture, including spectrum, sound wave reflection and holographic imaging data. The role of the multi-layer neural network in this link is to conduct deep learning and analysis of these data, so as to accurately identify the type of paint defects on the surface or interior of the furniture, and generate classification results including defect type and severity based on the characteristics of the defect.

[0122] A multi-layer neural network is a machine learning model that can extract deep features from complex data. In the present invention, the input of the neural network is feature data after multimodal data fusion, which contains information such as the color, thickness, surface texture and internal structure of the furniture surface paint. Through the multi-layer structure of the neural network, the model can automatically learn the characteristic patterns of different types of defects from these complex data, such as uneven color, surface cracks, abnormal thickness, bubbles, etc. During the training process of the neural network, the system uses a large amount of labeled furniture surface paint defect data for learning to form a classification model that can automatically identify new defects.

[0123] During the recognition process, the neural network first obtains multimodal feature data through the input layer, and then gradually extracts deep features from the data after processing through several hidden layers. The weights and biases of each layer are automatically adjusted by the model based on the training data. The neural network can learn the defect feature information hidden in different modal data. For example, the color difference in a certain area may correspond to the change in coating thickness, the surface texture anomaly in holographic imaging may represent paint cracks, and the specific amplitude pattern in the sound wave reflection data may indicate bubbles or cracks inside the furniture. Finally, after the output layer, the neural network classifies the data and generates a classification result containing the specific defect type and severity.

[0124] In actual operation, suppose the system detects that the color of a certain area on the surface of furniture is slightly different from the surrounding area through multimodal data fusion, and the acoustic wave data shows that there is a large amplitude change in this area, which may mean that the paint thickness is abnormal. By learning the characteristic patterns of this type of multimodal data, the neural network can determine that the area is a defect of uneven paint thickness, and further analyze the model to derive the severity of the defect (such as slight or severe thickness difference). Ultimately, the output of the neural network will be used to generate detailed defect classification data, including the type, location and severity of the defect.

[0125] The use of multi-layer neural networks for defect identification and classification has extremely high accuracy and adaptability. Since neural networks can automatically extract deep features from multimodal data and learn, they can adapt to various complex detection scenarios. Whether it is a high-gloss surface, furniture with complex structures, or areas with thick coatings, neural networks can accurately identify defects. In addition, neural networks can effectively process the association information between different types of defects. For example, surface color differences may be related to internal structural problems. Through multimodal data fusion and deep learning, neural networks can identify these potential associations and further improve the accuracy of detection.

[0126] In order to further optimize the recognition effect, a feedback enhancement mechanism is introduced in the present invention. The principle of the feedback enhancement mechanism is that in each round of defect recognition, the classification results generated by the neural network are fed back to the model to adjust the weights and parameters of the neural network. Specifically, the neural network will automatically adjust the weights through the feedback mechanism according to the errors in the recognition results (such as classification errors or low classification confidence), so that the model can more accurately capture subtle defect features in the next round of recognition. This self-adjustment process enables the model to continuously optimize its classification accuracy, especially when dealing with defects that are difficult to identify or have fuzzy boundaries. The feedback enhancement mechanism can significantly improve the reliability and accuracy of the classification results.

[0127] For example, suppose that during a certain recognition process, the neural network identifies a potential area of ​​abnormal coating thickness, but due to data noise or blurred edges, the confidence of the classification result is low. Through the feedback enhancement mechanism, the system will feed back this low-confidence result to the neural network, and the model will adjust the parameter settings for the next round based on this feedback, increase sensitivity to this type of problem, and further optimize the weight distribution, making subsequent classification results more accurate. Ultimately, after multiple rounds of feedback adjustment, the system can generate high-precision defect classification data and provide repair suggestions for each defect.

[0128] Generating a paint defect detection report is the final output of this process. Based on the defect classification data generated by the neural network, the system will comprehensively analyze the location, type and severity of the defects and generate a detailed paint defect detection report. The report not only contains a detailed description of each defect, but may also include repair suggestions to help operators handle and repair different types of paint problems accordingly. Through this data-driven automated report generation mechanism, the system can significantly improve detection efficiency and provide high-quality data support for the repair of paint defects on furniture surfaces.

[0129] like Figure 5 As shown, a system for implementing the furniture surface paint defect detection method, the system comprising:

[0130] The multimodal data acquisition module is used to obtain multimodal data of the furniture surface, and the multimodal data includes spectral data, acoustic wave reflection data and holographic imaging data collected by multi-frequency spectral sensors, acoustic wave sensors and holographic imaging sensors; the furniture surface paint defect detection process of the present invention is based on multimodal data acquisition and processing, and mainly uses multi-frequency spectral sensors, acoustic wave sensors and holographic imaging sensors. These sensors are used to obtain spectral data, acoustic wave reflection data and holographic imaging data on the furniture surface, and comprehensively scan the defects on the furniture surface and inside from multiple dimensions. Through this multimodal data acquisition, it can be ensured that different types of defects are detected, such as uneven surface color, abnormal thickness and internal structure problems.

[0131] The data processing module is used to perform preliminary processing on the multimodal data, including noise filtering and normalization processing, and generate preliminary multimodal fusion data; the data processing module first performs preliminary processing on the collected multimodal data, including noise filtering and normalization processing, to eliminate environmental interference and data deviation. This step generates preliminary multimodal fusion data as the basis for subsequent analysis.

[0132] The spatiotemporal characteristic fusion module is used to perform spatiotemporal characteristic fusion processing based on the preliminary multimodal fusion data, and dynamically adjust the weights of different modal data using a multimodal adaptive weight allocation algorithm to generate fused multimodal feature data; the spatiotemporal characteristic fusion module uses a multimodal adaptive weight allocation algorithm to dynamically adjust the weight of each type of data according to the characteristics of different modal data to generate fused multimodal feature data. This weight adjustment ensures that the data can complement each other in complex surfaces or multiple detection environments to maximize recognition accuracy.

[0133] The defect feature extraction module is used to perform multi-layer progressive defect feature extraction based on the fused multi-modal feature data using a progressive residual self-correction algorithm to generate verified defect feature data; the defect feature extraction module uses a progressive residual self-correction algorithm to perform multiple rounds of iterative correction on the fused feature data to extract multi-layer progressive defect feature data. In this way, the system can gradually optimize the model parameters and enhance the detection accuracy of surface and internal defects, especially in the case of complex surfaces or hidden defects.

[0134] The feedback and dynamic adjustment module is used to perform adaptive feedback and dynamic adjustment based on the verified defect feature data, adjust the parameters of the multi-frequency spectrum sensor, acoustic wave sensor and holographic imaging sensor through feedback residual analysis, and perform a second data acquisition to generate the second multimodal fusion data; the feedback and dynamic adjustment module analyzes the residual and adaptively adjusts the parameters of the sensor based on the extracted defect feature data, and then performs a second data acquisition. This mechanism ensures that the system can optimize the sensor according to the error of the initial acquisition, further improving the accuracy of the data.

[0135] The defect recognition and classification module is used to use a multi-layer neural network to perform defect recognition and classification based on the second multi-modal fusion data, generate defect classification data including defect type and severity, and generate a paint defect detection report based on the defect classification data. The defect recognition and classification module classifies the multi-modal data of the second data collection through a multi-layer neural network, identifies the defect type and its severity, and generates a detailed paint defect detection report. This report provides operators with an accurate description of the location, type and severity of defects on the surface and inside of the furniture, which is convenient for subsequent repairs and quality control.

[0136] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0137] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for detecting paint defects on furniture surfaces, characterized in that: The following steps are involved: Acquire multimodal data of the furniture surface, wherein the multimodal data includes spectral data, sound wave reflection data and holographic imaging data, and generate preliminary multimodal fusion data; Based on the preliminary multimodal fusion data, performing spatiotemporal characteristic fusion processing, dynamically adjusting the weights of different modal data using a multimodal adaptive weight allocation algorithm, and generating fused multimodal feature data; Based on the fused multimodal feature data, a progressive residual self-correction algorithm is used to perform multi-layer progressive defect feature extraction to generate verified defect feature data; the progressive residual self-correction algorithm optimizes the accuracy of defect feature extraction through multiple rounds of correction; Based on the verified defect feature data, adaptive feedback and dynamic adjustment are performed, and the sensor parameters for collecting multimodal data are adjusted through feedback residual analysis, and a second data collection is performed to generate second multimodal fusion data; in the adaptive feedback and dynamic adjustment step, the collection parameters of the multi-frequency spectrum sensor, the acoustic wave sensor and the holographic imaging sensor are dynamically adjusted through feedback residual analysis, and the specific adjustment includes: Adjust the wavelength range of the multi-frequency spectral sensor to capture surface color changes; Adjust the amplitude and frequency of the acoustic sensor to detect defects inside furniture; Adjusting the resolution of holographic imaging sensors to improve the detection accuracy of surface texture and structure; Based on the second multimodal fusion data, a multi-layer neural network is used to perform defect recognition and classification, to generate defect classification data including defect type and severity, and to generate a paint defect detection report based on the defect classification data.

2. The method for detecting paint defects on furniture surface according to claim 1, characterized in that: The preliminary processing of the multimodal data includes noise filtering of spectral data, acoustic reflection data and holographic imaging data, wherein the filtering step uses a preset noise threshold value to remove environmental interference and sensor errors; and normalizing the data of different modalities through a normalization processing algorithm so that the data of each modality can be compared and fused within a unified numerical range, thereby generating preliminary multimodal fusion data.

3. The method for detecting paint defects on furniture surface according to claim 1, characterized in that: The multimodal adaptive weight allocation algorithm dynamically adjusts the weights of different modal data through the following formula: in, For the The weight of the modal data; For the The accuracy of the modal data; is the total number of modal data.

4. The method for detecting paint defects on furniture surface according to claim 3, characterized in that: The accuracy Calculated by the following formula: in, For the The accuracy of the modal data; For the The number of sampling points of each modal data; For the The modal data is in The value of the data point; For the The average value of all modal data points.

5. The method for detecting paint defects on furniture surface according to claim 1, characterized in that: The progressive residual self-correction algorithm corrects the residual in defect feature extraction by the following formula: in, is the current corrected residual value; is the residual value of the previous round; is the preset learning rate; The value extracted for the current feature.

6. The method for detecting paint defects on furniture surface according to claim 5, characterized in that: The progressive residual self-correction algorithm optimizes the accuracy of defect feature extraction through multiple rounds of correction, including the following steps: in each round of residual correction, the extracted defect features are analyzed, and the residuals are fed back to the feature extraction model, the model parameters are adjusted, and finally optimized defect feature data is generated. The defect feature data is used to identify the defect location and severity on the surface and inside of the furniture.

7. The method for detecting paint defects on furniture surfaces according to claim 1, characterized in that: The second data acquisition is performed based on the dynamically adjusted sensor parameters to generate second multimodal fusion data, and the second multimodal fusion data is used to improve the accuracy of defect detection on complex surfaces and inside furniture.

8. The method for detecting paint defects on furniture surface according to claim 1, characterized in that: The multi-layer neural network adjusts the parameters and weights of the model through a feedback enhancement mechanism, and the feedback enhancement mechanism includes the following steps: in each round of defect identification, the identified defect classification results are fed back to the neural network, and the weights of the neural network are adjusted through the feedback to optimize the classification accuracy, and finally generate defect classification data and repair suggestions.

9. A system for implementing the method for detecting paint defects on furniture surfaces according to any one of claims 1 to 8, characterized in that: The system comprises: A multimodal data acquisition module, used to acquire multimodal data of the furniture surface, wherein the multimodal data includes spectral data, acoustic wave reflection data and holographic imaging data acquired by a multi-frequency spectral sensor, an acoustic wave sensor and a holographic imaging sensor; A data processing module, used to perform preliminary processing on the multimodal data, including noise filtering and normalization processing, and generate preliminary multimodal fusion data; A spatiotemporal characteristic fusion module, for performing spatiotemporal characteristic fusion processing based on the preliminary multimodal fusion data, dynamically adjusting the weights of different modal data using a multimodal adaptive weight allocation algorithm, and generating fused multimodal feature data; A defect feature extraction module, used to perform multi-layer progressive defect feature extraction based on the fused multi-modal feature data using a progressive residual self-correction algorithm to generate verified defect feature data; A feedback and dynamic adjustment module, which is used to perform adaptive feedback and dynamic adjustment based on the verified defect feature data, adjust the parameters of the multi-frequency spectrum sensor, the acoustic wave sensor and the holographic imaging sensor through feedback residual analysis, and perform a second data acquisition to generate second multimodal fusion data; A defect recognition and classification module is used to perform defect recognition and classification based on the second multimodal fusion data using a multi-layer neural network, generate defect classification data including defect type and severity, and generate a paint defect detection report based on the defect classification data.

Images