Image processing method for evaluating the polishing effect of a housing

Abstract

The application discloses an image processing method for evaluating polishing effect of a shell, and belongs to the technical field of image processing, and specifically comprises the following steps: collecting a dynamic thermal imaging sequence of a shell surface and extracting a temperature response curve, and dividing a heat conduction consistency area and a heat anomaly area; obtaining a set of thermal physical parameters by analyzing the temperature change law of the heat conduction consistency area; synchronously collecting a polarized scattered light image and extracting polarized image features; performing fusion processing on the set of thermal physical parameters, the spatial distribution information of the heat anomaly area and the polarized image features to generate a comprehensive feature representation; finally, the matching degree with a standard feature template is calculated through a hierarchical model to determine the polishing quality grade and generate an evaluation report; and the application realizes multi-modal precise evaluation of polishing quality and reliable identification of local defects.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an image processing method for evaluating the polishing effect of a housing. Background Technology

[0002] The quality of surface polishing is a key indicator of precision manufacturing process, directly affecting the product's appearance, corrosion resistance, and lifespan. In fields such as consumer electronics, automotive parts, and high-end equipment manufacturing, achieving objective and accurate evaluation of polishing results is crucial for process optimization and quality control.

[0003] Currently, various surface quality inspection methods have been developed in this field. Mainstream non-contact inspection technologies are primarily based on optical imaging principles, evaluating polishing effects by analyzing the surface's light reflection characteristics. These methods typically use high-resolution cameras to acquire images of the casing surface and establish an evaluation system by calculating parameters such as grayscale distribution and texture features in the images. Another common approach is to utilize laser scanning technology to obtain three-dimensional surface topography data, achieving quantitative assessment by analyzing height differences and roughness parameters.

[0004] Existing technologies typically rely on a single type of data acquisition method, such as visible light imaging or three-dimensional topographic scanning. This approach struggles to effectively reconcile the combined effects of different physical properties on the inspection results when inspecting workpieces with complex surface characteristics or localized minor defects. Furthermore, existing methods often treat the surface as a homogeneous whole during data processing, lacking in-depth analysis of the differences in characteristics across different regions. This results in limitations in simultaneously assessing overall quality and locating local anomalies. These limitations have prompted the industry to seek novel inspection solutions that can integrate multi-dimensional information and consider both overall and local features. Summary of the Invention

[0005] The purpose of this invention is to provide an image processing method for evaluating the polishing effect of a housing, thereby solving the following technical problems:

[0006] Existing methods for evaluating the quality of surface polishing mainly rely on a single type of test data, which makes it difficult to effectively distinguish the mixed influence of surface micro-geometry and material properties, and lacks accurate identification of local abnormal areas.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An image processing method for evaluating the polishing effect of a housing includes the following steps:

[0009] S1. Acquire dynamic thermal imaging sequences of the outer shell surface under controllable thermal excitation conditions, including basic state images before thermal excitation begins, temperature change image sequences during thermal excitation, and temperature recovery images after thermal excitation stops.

[0010] S2. Extract the response curve of temperature change over time for each pixel in the dynamic thermal imaging sequence, and divide the outer shell surface into a region of consistent thermal conduction and a region of thermal anomaly based on the morphological characteristics of the response curve.

[0011] S3. Analyze the temperature change pattern in the heat conduction uniformity region, calculate the surface temperature rise gradient, surface temperature distribution uniformity index and heat diffusion time constant, and combine them to form a set of thermophysical parameters.

[0012] S4. During the thermal excitation process, several acquisition times are set to synchronously acquire polarized scattered light images of the outer shell surface and extract polarization image features.

[0013] S5. The set of thermophysical parameters, the spatial distribution information of thermal anomaly regions, and polarization image features are fused together to generate a comprehensive feature representation of the surface polishing state.

[0014] S6. Input the comprehensive feature representation into the grading model, calculate the matching degree with the standard feature template of each polishing quality level, determine the polishing quality level based on the matching degree, and generate an evaluation report.

[0015] As a further aspect of the present invention: in step S2, the specific process of dividing the outer shell surface into a thermal conductivity consistency region and a thermal anomaly region based on the morphological characteristics of the response curve is as follows:

[0016] For each pixel location in the dynamic thermal imaging sequence, complete temperature data from the start of thermal excitation to the end of temperature recovery is extracted to form a temperature time series. Multimodal curve fitting is performed on the temperature time series to obtain a temperature change curve containing multiple components. The first derivative sequence of the temperature change curve is calculated to obtain the temperature change rate at each moment. The global maximum and global minimum points are located on the temperature change rate curve. The amplitude difference between the maximum and minimum points is calculated to determine the zero-crossing point where the temperature change rate curve changes from a positive value to a negative value. The time offset of the zero-crossing point relative to the moment when thermal excitation stops is recorded. Pixels with amplitude differences within a predetermined range and time offsets that meet predetermined standards are classified as thermal conduction consistency regions.

[0017] Pixels whose amplitude difference exceeds the predetermined range or whose time offset does not meet the predetermined standard are classified as thermal anomalous pixels. Spatial connectivity analysis is performed on thermal anomalous pixels to merge adjacent thermal anomalous pixels into continuous thermal anomalous regions.

[0018] As a further aspect of the present invention: the specific process of performing multimodal curve fitting on the temperature time series is as follows:

[0019] A multimodal fitting equation is established that includes heat conduction, heat convection and heat radiation components. The heat conduction component adopts the form of an exponential decay function, the heat convection component adopts the form of a linear function, and the heat radiation component adopts the form of a power function.

[0020] The least squares method is used to solve the coefficient parameters of each component in the multimodal fitting equation. Based on the coefficient parameters, the contribution of each component to the overall temperature response is calculated. The component with the largest contribution is determined as the dominant thermal response mode. The pixel region is initially classified based on the type of dominant thermal response mode. The region dominated by the thermal conduction component is marked as the first type region, the region dominated by the thermal convection component is marked as the second type region, and the region dominated by the thermal radiation component is marked as the third type region.

[0021] As a further aspect of the present invention: In step S3, the specific process of analyzing the temperature change law in the heat conduction uniformity region and calculating the surface temperature rise gradient, surface temperature distribution uniformity index, and heat diffusion time constant is as follows:

[0022] Determine the start time of thermal excitation and the time when the temperature reaches stability, and calculate the average rate of change of surface temperature within this time interval as the surface temperature rise gradient.

[0023] During the temperature stabilization phase of the thermal excitation process, the standard deviation of the temperature values ​​of all pixels within the heat conduction uniformity region is calculated as an indicator of the surface temperature distribution uniformity.

[0024] During the temperature recovery phase after thermal excitation stops, the temperature drop curve is fitted with an exponential function, and the thermal diffusion time constant is extracted from the coefficients of the exponential term of the fitted function. The surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant are normalized in terms of dimensions and combined to form a set of thermophysical parameters.

[0025] As a further aspect of the present invention: In step S4, several acquisition times are set during the thermal excitation process to simultaneously acquire polarized scattered light images of the outer shell surface, and the specific process of extracting polarization image features is as follows:

[0026] Based on the analysis of the temperature change image sequence during thermal excitation, the moments when the temperature change rate reaches its peak and when the temperature distribution pattern undergoes abrupt changes are determined. These moments are marked as acquisition moments. At each acquisition moment, a polarization light source and a polarization camera are triggered to acquire data synchronously. The polarization light source provides incident light with a specific polarization direction, and the polarization camera is equipped with a rotatable analysis mirror. At each acquisition moment, a set of polarized scattered light images is acquired. The image set contains multiple scattered light images under different polarization directions of the analysis mirror. The polarization modulation depth distribution map and depolarization degree distribution map are extracted from the polarized scattered light images as polarization image features.

[0027] As a further aspect of the present invention: In step S5, the specific process of fusing the set of thermophysical parameters, the spatial distribution information of the thermal anomaly region, and the polarization image features to generate a comprehensive feature representation of the surface polishing state is as follows:

[0028] The surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant in the thermophysical parameter set are mapped to three independent thermal feature channels, and the three thermal feature channels are combined into a thermal feature cube. The spatial distribution information of the thermal anomaly region is converted into a spatial weighted distribution map, and the spatial weighted distribution map is used to perform region-selective enhancement of the polarization modulation depth distribution map and the depolarization degree distribution map.

[0029] The enhanced polarization modulation depth distribution map and depolarization degree distribution map are concatenated with the thermal feature cube at the channel level to form a multimodal feature cube. A three-dimensional convolution operation is performed on the multimodal feature cube to extract the joint features of the spatial-feature dimensions. The joint features are flattened into a one-dimensional vector as a comprehensive feature representation of the surface polishing state.

[0030] As a further aspect of the present invention: the specific process of using the spatial weight distribution map to perform region-selective enhancement of the polarization modulation depth distribution map and the depolarization degree distribution map is as follows:

[0031] The spatial weight distribution map is multiplied by the surface temperature distribution uniformity index channel in the thermal feature cube to obtain the enhanced weight distribution map. The enhanced weight distribution map is then subjected to a nonlinear transformation. The transformed enhanced weight distribution map is then multiplied pixel by pixel with the polarization modulation depth distribution map and the depolarization degree distribution map to obtain the region-enhanced polarization modulation depth distribution map and the region-enhanced depolarization degree distribution map.

[0032] The correlation coefficient matrix between the polarization modulation depth distribution map and the depolarization degree distribution map of the region enhancement is calculated. The correlation coefficient matrix is ​​then used to perform feature interaction with the thermal diffusion time constant channel in the thermal feature cube to generate a polarization feature representation with thermal-optical coupling characteristics.

[0033] As a further aspect of the present invention: In step S6, the specific process of inputting the comprehensive feature representation into the hierarchical model and calculating the matching degree with the standard feature templates of each polishing quality level is as follows:

[0034] A standard feature template library for polishing quality grades is constructed. Each template corresponds to a comprehensive feature representation of a polishing quality grade. The comprehensive feature representation of the shell to be tested is compared with each standard feature template in the template library to calculate the Euclidean distance in the feature space. The Euclidean distance is converted into a matching score. The matching score is negatively correlated with the Euclidean distance. The polishing quality grade with the highest matching score is selected as the final evaluation result, and an evaluation report is generated. The evaluation report includes the polishing quality grade number and the matching score of each standard feature template.

[0035] The beneficial effects of this invention are:

[0036] This invention constructs a collaborative analysis system of thermophysical parameter sets and polarization image features by fusing multimodal data from dynamic thermal imaging sequences and polarized scattered light images. This overcomes the limitation of traditional single optical detection methods, which struggle to distinguish the mixed influence of surface micro-geometry and material thermophysical properties. By establishing an automatic division mechanism between thermal conductivity consistency regions and thermal anomaly regions, and innovatively utilizing the spatial distribution information of thermal anomaly regions to generate a weighted distribution map, it achieves region-selective enhancement of polarization image features. This effectively solves the problem of traditional methods treating the surface as a uniform whole and lacking the ability to accurately identify local anomalies. Based on the deep fusion of thermal feature cubes and enhanced polarization features, a comprehensive feature representation with thermal-optical coupling characteristics is generated. Combined with the standard feature template matching mechanism of the hierarchical model, this significantly improves the accuracy and reliability of shell polishing quality evaluation, providing more comprehensive and accurate technical support for process optimization in the field of precision manufacturing. Attached Figure Description

[0037] The invention will now be further described with reference to the accompanying drawings.

[0038] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 As shown, the present invention is an image processing method for evaluating the polishing effect of a housing, comprising the following steps:

[0041] S1. Acquire dynamic thermal imaging sequences of the outer shell surface under controllable thermal excitation conditions, including basic state images before thermal excitation begins, temperature change image sequences during thermal excitation, and temperature recovery images after thermal excitation stops.

[0042] S2. Extract the response curve of temperature change over time for each pixel in the dynamic thermal imaging sequence, and divide the outer shell surface into a region of consistent thermal conduction and a region of thermal anomaly based on the morphological characteristics of the response curve.

[0043] S3. Analyze the temperature change pattern in the heat conduction uniformity region, calculate the surface temperature rise gradient, surface temperature distribution uniformity index and heat diffusion time constant, and combine them to form a set of thermophysical parameters.

[0044] S4. During the thermal excitation process, several acquisition times are set to synchronously acquire polarized scattered light images of the outer shell surface and extract polarization image features.

[0045] S5. The set of thermophysical parameters, the spatial distribution information of thermal anomaly regions, and polarization image features are fused together to generate a comprehensive feature representation of the surface polishing state.

[0046] S6. Input the comprehensive feature representation into the grading model, calculate the matching degree with the standard feature template of each polishing quality level, determine the polishing quality level based on the matching degree, and generate an evaluation report.

[0047] In a preferred embodiment of the present invention, the specific process of dividing the outer shell surface into a thermally consistent region and a thermally abnormal region based on the morphological characteristics of the response curve in step S2 is as follows:

[0048] For temperature time series extraction at each pixel location, the dynamic thermal imaging sequence must first be preprocessed. Preprocessing includes image denoising and pixel coordinate calibration. Image denoising employs a combination of Gaussian and median filtering to remove random noise while preserving temperature abrupt changes. Pixel coordinate calibration is achieved through image registration technology to eliminate pixel position shifts caused by slight equipment vibrations during thermal imaging, ensuring that temperature data at the same physical location corresponds to the same pixel coordinates in different frames. After preprocessing, the temperature values ​​are extracted from each frame frame one by one according to pixel coordinates and arranged chronologically to form the temperature time series for that pixel. The time interval of the series is consistent with the frame rate of the thermal imager.

[0049] Multimodal curve fitting is a key step in analyzing the physical implications of temperature time series. Its core is constructing a fitting equation that includes three components: heat conduction, heat convection, and heat radiation. The heat conduction component uses an exponentially decaying function because the heat conduction process inside a solid follows Fourier's law, with temperature changes exhibiting an exponential decay over time, accurately reflecting the heat transfer pattern within the shell material. The heat convection component uses a linear function, corresponding to the convective heat transfer process between the shell surface and the surrounding air. The temperature change rate in this process is relatively stable, and the linear function effectively characterizes its contribution to the overall temperature response. The heat radiation component uses a power function, consistent with the physical law that heat radiation is proportional to the fourth power of temperature, suitable for describing the radiative heat dissipation process from the shell surface to the environment.

[0050] The fitted equation is solved using the least squares method. Before solving, the temperature time series needs to be normalized, converting the temperature values ​​to the range of 0 to 1 to eliminate the influence of magnitude differences on the coefficient solution. During the solution process, initial coefficient values ​​for each component are preset based on heat transfer theory. For example, the attenuation coefficient of the heat conduction component is estimated based on the thermal conductivity of the shell material, and the linear coefficient of the heat convection component is estimated based on the ambient wind speed. Then, the coefficient parameters are continuously adjusted through iterative calculations to minimize the sum of squared residuals between the fitted curve and the original temperature time series. To improve the solution accuracy, the gradient descent method is used to optimize the iteration process. By calculating the partial derivatives of the residual function with respect to each coefficient, the adjustment direction and step size of the coefficients are determined, accelerating the iteration convergence speed.

[0051] The contribution of each component is calculated using variance contribution as an indicator. By decomposing the total variance of the temperature time series, the proportion of variance explained by each component is obtained, and this proportion is the contribution of the corresponding component. The component with the largest contribution is identified as the dominant thermal response mode. Based on this, the pixel regions are initially classified as follows: the first type of region dominated by the thermal conduction component indicates that heat transfer in this region is mainly through internal conduction within the solid and is less affected by environmental factors; the second type of region dominated by the thermal convection component indicates that the convective heat transfer between this region and the environment is significant, and it is usually located in the ventilation holes or surface protrusions of the shell; the third type of region dominated by the thermal radiation component indicates that its heat loss mainly depends on radiation, and it mostly corresponds to areas with high surface smoothness of the shell. The preliminary classification results provide a physical basis for subsequent consistency judgment, making the region division more consistent with the actual heat transfer characteristics.

[0052] The rate of temperature change is calculated by taking the first derivative of the fitted temperature change curve. The central difference method is used to calculate the derivative sequence. This method calculates the rate of change at the intermediate moment using temperature values ​​from three adjacent moments, offering higher accuracy compared to forward or backward differencing and accurately capturing the instantaneous characteristics of temperature changes. To locate the global maximum and minimum points on the temperature change rate curve, a sliding window method is used to traverse the entire derivative sequence. The window size is determined based on the thermal imager's frame rate, typically set to 5 to 10 data points. By comparing the extreme values ​​within the window with the global extreme values, misjudgments caused by local noise are avoided. The global maximum point corresponds to the moment of the fastest temperature increase, and the global minimum point corresponds to the moment of the fastest temperature decrease. The difference between these two values ​​reflects the dynamic range of the temperature response at that pixel location.

[0053] Locating the zero-crossing point is crucial for determining the phase characteristics of the thermal response. The zero-crossing point is the moment when the rate of temperature change changes from positive to negative, corresponding to the peak point of the temperature time series. Physically, it represents the critical point where the temperature at that location in the shell transitions from the rising phase to the falling phase. The time offset of the zero-crossing point relative to the moment when thermal excitation stops reflects the hysteresis characteristics of the thermal response. The smaller the offset, the faster the thermal response in that region and the higher the heat transfer efficiency.

[0054] The classification of regions with consistent thermal conductivity requires setting predetermined ranges for amplitude differences and predetermined standards for time offsets. These two parameters are determined through statistical analysis of extensive experimental data. A standard sample with the same material and uniform structure as the shell is selected for thermal response experiments. The amplitude differences and time offsets of each pixel on its surface are collected, and the statistical distribution range of these parameters is calculated. The mean plus or minus two standard deviations is used as the predetermined range and standard. Pixels with amplitude differences within this range and time offsets conforming to the standard are classified as regions with consistent thermal conductivity. These regions exhibit stable and consistent thermal transfer characteristics, uniform material structure, and no obvious defects.

[0055] The classification criteria for thermal anomaly pixels are those with amplitude differences exceeding a predetermined range or time offsets not meeting predetermined standards. These pixels exhibit significantly different thermal response characteristics compared to their surrounding areas, potentially corresponding to structural defects or material inhomogeneities in the shell. Spatial connectivity analysis of thermal anomaly pixels employs an 8-neighborhood connectivity judgment method, determining whether adjacent pixels in the vertical, horizontal, and four diagonal directions of an anomaly pixel are also anomalies. If adjacent anomalies exist, they are grouped into the same connected region. By scanning the entire thermal imaging image and traversing all thermal anomaly pixels, connected regions are marked and merged to form continuous thermal anomaly regions. To eliminate interference from isolated noise points, the merged connected regions are filtered by area, removing isolated connected regions with areas smaller than a set threshold. The resulting thermal anomaly regions accurately reflect abnormal heat conduction locations on the shell surface, providing precise localization for subsequent defect detection.

[0056] In another preferred embodiment of the present invention, the specific process of analyzing the temperature change law of the heat conduction uniformity region and calculating the surface temperature rise gradient, surface temperature distribution uniformity index, and heat diffusion time constant in step S3 is as follows:

[0057] The primary task in calculating the surface temperature gradient is to determine the start and stabilization points of thermal excitation. The start of thermal excitation is determined through synchronous analysis of the trigger signal and the temperature sequence. Using the timestamp of the thermal excitation device's activation signal as a reference, the first moment in the temperature sequence where the temperature begins to rise continuously from the environmental baseline is matched; this moment is the start of thermal excitation. The stabilization point is determined by monitoring the rate of temperature change. A very small threshold for the rate of temperature change is set. When the rate of change of multiple consecutive data points in the temperature time series is below this threshold, the temperature is considered to have entered a stable phase, and the start moment of this consecutive sequence is taken as the stabilization point. After determining these two key moments, the total temperature change within this time interval is calculated, i.e., the difference between the average temperature at the stabilization point and the initial temperature at the start of thermal excitation. Dividing the total temperature change by the time interval between the two moments yields the average rate of change, which is the surface temperature gradient. This indicator directly reflects the material's temperature response speed under thermal excitation.

[0058] The calculation of the surface temperature distribution uniformity index focuses on the temperature stabilization phase of the thermal excitation process. First, image frames corresponding to the stabilization phase are extracted from the dynamic thermal imaging sequence, and the temperature values ​​of all pixels within the heat conduction consistency region are extracted to form a temperature dataset. To avoid interference from isolated outliers, this dataset needs to be preprocessed, using the 3σ criterion to remove extreme temperature values ​​exceeding the mean plus or minus three standard deviations. The standard deviation is calculated for the processed temperature data. The magnitude of the standard deviation directly reflects the dispersion of temperature distribution within the region; a smaller standard deviation indicates a more uniform surface temperature distribution. Therefore, this standard deviation is directly used as the surface temperature distribution uniformity index, which reflects the spatial consistency of material heat conduction.

[0059] The extraction of the thermal diffusion time constant targets the temperature recovery phase after the thermal excitation stops. First, the moment when the thermal excitation stops is determined, using the timestamp of the thermal excitation device's shutdown signal as a basis, and matching the moment in the temperature sequence when the temperature begins to decrease from a stable state. Temperature decrease data after this moment is extracted from the temperature time series as the raw data for the temperature decrease curve. Considering that the temperature recovery process follows the exponential decay law of thermal diffusion, a single exponential function is used to fit the temperature decrease curve. The fitting process employs the least squares method, and the coefficients of the fitting function are determined through iterative optimization. The thermal diffusion time constant is directly related to the coefficient of the exponential term of the fitting function; the reciprocal of this coefficient is the thermal diffusion time constant, reflecting the rate characteristics of heat diffusion within the material.

[0060] Dimensional normalization is a crucial step in eliminating differences in the magnitudes of various parameters. Since the dimensions of the surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant are different, direct combination can lead to the dominance of parameters with larger magnitudes in subsequent analyses. A linear normalization method is used to process each parameter, mapping its original value to the interval between 0 and 1. The mapping process is based on the maximum and minimum values ​​of the parameter in a standard sample library, ensuring the comparability of the normalized parameter values. The three normalized parameters are then combined in a fixed order to form a set of thermophysical parameters encompassing the material's thermal conductivity, spatial consistency, and diffusion characteristics, providing standardized input for subsequent material property analysis.

[0061] In another preferred embodiment of the present invention, in step S4, the specific process of setting several acquisition times during the thermal excitation process, synchronously acquiring polarized scattered light images of the outer shell surface, and extracting polarization image features is as follows:

[0062] The determination of the acquisition time needs to be based on an in-depth analysis of the temperature change pattern during the thermal excitation process. Based on the temperature change rate curve obtained in S2, the moment when the temperature change rate reaches its peak is located. This moment corresponds to the stage where the material surface temperature rises the fastest, and the change in the internal thermal stress distribution of the material is most significant at this time. Simultaneously, by monitoring the dynamic changes of the surface temperature distribution uniformity index in the heat conduction consistency region, when this index shows a sudden change, it indicates a change in the temperature distribution pattern, and this sudden change moment is also marked as a critical acquisition moment. To ensure comprehensive acquisition, auxiliary acquisition moments are added at fixed time intervals between the start of thermal excitation, the temperature stabilization moment, and the aforementioned critical moments, forming a sequence of acquisition moments covering the entire thermal excitation process. Each acquisition moment is precisely marked with a timestamp.

[0063] Synchronous acquisition of polarized scattered light images relies on a multi-device collaborative control mechanism. The system comprises a synchronization trigger module centered on a main controller. The main controller sends synchronization trigger signals to the polarization light source and polarization camera according to a preset acquisition time sequence. The polarization light source uses a narrow-band polarized LED, providing linearly polarized incident light in a specific polarization direction. This polarization direction can be preset and fixed according to experimental requirements. The polarization camera is equipped with a high-resolution CCD sensor and a rotatable analyzer. The rotation angle of the analyzer is precisely controlled by a stepper motor. At each acquisition time, the stepper motor drives the analyzer to rotate sequentially to multiple preset polarization directions. The polarization camera completes one image acquisition in each polarization direction, forming an image set containing multiple images with different polarization directions. During acquisition, it is crucial to ensure that the incident angle of the polarization light source and the shooting angle of the polarization camera remain fixed to avoid interference from changes in geometric parameters on the scattered light signal.

[0064] The extraction of polarization image features focuses on two core indicators: polarization modulation depth distribution map and depolarization degree distribution map. Polarization modulation depth reflects the ability of a material surface to modulate incident polarized light, and its calculation is based on the image grayscale values ​​under different polarization directions at the same acquisition time. A sine curve is fitted to the grayscale values ​​of each pixel in the image group under different polarization directions; the ratio of the amplitude of the fitted curve to the average value is the polarization modulation depth of that pixel. This process is repeated for all pixels to generate the polarization modulation depth distribution map. Depolarization degree reflects the degree of change in the polarization state of scattered light. It is calculated by dividing the difference between the maximum and minimum grayscale values ​​in the image group by the sum of the maximum and minimum grayscale values, thus obtaining the depolarization degree of each pixel and generating the depolarization degree distribution map.

[0065] Image preprocessing is required during feature extraction to improve feature quality. First, dark current correction and background subtraction are performed on the acquired polarized scattered light images to eliminate camera noise and ambient light interference. Gaussian filtering is then used to remove high-frequency noise while preserving detailed polarization features. Image registration techniques are employed to eliminate positional shifts caused by minor device jitter between images of different polarization directions, ensuring that the grayscale value of the same pixel accurately corresponds to the scattered light signal in different polarization directions. The extracted polarization modulation depth distribution map and depolarization degree distribution map need to be timestamped with the temperature data at the corresponding acquisition time to form a "temperature state-polarization feature" associated dataset, providing data support for subsequently establishing the mapping relationship between thermophysical and optical properties.

[0066] In another preferred embodiment of the present invention, the specific process of fusing the set of thermophysical parameters, the spatial distribution information of the thermal anomaly region, and the polarization image features in step S5 to generate a comprehensive feature representation of the surface polishing state is as follows:

[0067] The construction of thermal feature channels is based on dimensionally normalized thermophysical parameters. The surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant are all single-valued parameters, which need to be spatially expanded into single-channel feature maps with dimensions consistent with the polarization image. Spatial expansion employs an interpolation mapping method, using the spatial range of the thermal conductivity consistency region as a benchmark, uniformly mapping the single-valued parameters to each pixel position within that region, while thermal anomaly regions are filled with preset low-weight values. Each of the three parameters generates an independent thermal feature channel through this process, with the pixel size and spatial resolution of each channel perfectly matching the polarization image to ensure spatial consistency during subsequent fusion. The three thermal feature channels are then superimposed dimensionally to form a thermal feature cube with a height equal to the number of feature channels and a width and height consistent with the image, thus completing the spatial representation of the thermophysical parameters.

[0068] The conversion of spatial distribution information of thermal anomaly regions into a spatial weighted distribution map requires two steps: binarization and weight assignment. First, the spatial distribution results of the thermal anomaly regions are converted into a binary image, with pixel values ​​in the thermal anomaly regions set to 0 and pixel values ​​in thermal conductivity consistency regions set to 1, forming an initial spatial weighted map. To highlight the reliability differences at different locations within the thermal conductivity consistency region, a distance-weighted mechanism is introduced. Pixels within the thermal conductivity consistency region are assigned different weights based on their distance to the nearest thermal anomaly region; the farther the distance, the larger the weight. This constructs a spatial weighted distribution map containing spatial confidence information. This map quantifies the feature reliability of different regions, providing a basis for subsequent selective enhancement of polarization features.

[0069] The core of region-selective enhancement is establishing a spatial correlation adjustment mechanism between thermal and polarization features. First, a dot product operation is performed between the spatial weight distribution map and the surface temperature distribution uniformity index channel in the thermal feature cube. This operation deeply couples the spatial weights with the temperature uniformity information; regions with more uniform temperature distribution receive amplified weights, while regions with slight temperature fluctuations but still within a consistent range have appropriately reduced weights. The final output is an enhanced weight distribution map. To avoid feature distortion due to oversaturation of weights, a nonlinear transformation is performed on the enhanced weight distribution map, using the Sigmoid function to compress the weights to the range of 0.1 to 0.9. This preserves the weight differences between regions while preventing excessive suppression of features by extreme weights.

[0070] The transformed enhanced weight distribution map is multiplied pixel-by-pixel with the polarization modulation depth distribution map and the depolarization degree distribution map to achieve regional enhancement of polarization features. During this process, the polarization features of regions with good thermal conductivity and uniform temperature distribution are significantly enhanced, while the polarization features of thermal anomaly regions and regions with poor consistency are weakened, ensuring that the subsequent fusion process focuses on the effective feature regions reflecting the polishing state. The two enhanced polarization feature maps undergo edge smoothing, and mean filtering is used to eliminate local noise introduced by pixel-by-pixel multiplication, maintaining the spatial continuity of the features.

[0071] The correlation coefficient matrix is ​​calculated to uncover the intrinsic relationship between two enhanced polarization features. Using pixel blocks as units, the enhanced polarization modulation depth distribution map and depolarization degree distribution map are divided into non-overlapping local regions. The correlation coefficient between the two features within each local region is calculated, forming a local correlation coefficient matrix. The size of the matrix corresponds to the number of local regions. The feature interaction process is implemented using matrix element-wise multiplication. The local correlation coefficient matrix is ​​multiplied element-wise by the thermal diffusion time constant channel in the thermal feature cube. The thermal diffusion time constant reflects the thermal conductivity of the material. Its combination with the polarization feature correlation enhances the thermal-optical coupling characteristics, generating a polarization feature representation that combines thermal conductivity information and polarization correlation information. This feature representation retains the details of optical properties while incorporating the constraints of the material's thermophysical properties.

[0072] The multimodal feature cube is constructed through channel-level stitching. The enhanced polarization modulation depth distribution map, depolarization degree distribution map, and polarization feature representation after feature interaction are superimposed on the original thermal feature cube along the channel dimension. At this point, the thermal feature cube provides 3 channels, and the polarization-related features provide 3 channels, together forming a 6-channel multimodal feature cube. The spatial dimension of this cube is consistent with the original image, and the channel dimension includes the core features of both modalities. Before stitching, standard deviation normalization is performed on all feature channels to ensure that the numerical ranges of different modal features are on the same order of magnitude, preventing any one type of feature from dominating the subsequent convolution process.

[0073] 3D convolution is crucial for extracting joint spatial-feature information. A 3×3×3 convolution kernel is used, with the number of kernels determined by the feature extraction requirements. A stride of 1 is set to preserve complete spatial information. During convolution, each kernel simultaneously slides across both the spatial and channel dimensions of the feature cube, capturing the spatial adjacency relationships of pixels and the correlation information between different modal features, thus achieving the transformation from independent modal features to fused modal features. After multiple layers of 3D convolution, the output feature map retains its 3D structure, but its channel dimension features now contain the joint semantics of space and modality.

[0074] The flattening process for joint features must maintain the consistency of feature order, converting the feature map output from the 3D convolution into a one-dimensional vector in channel-first order. The flattened vector dimension is the product of the number of channels, height, and width of the feature map. This vector integrates quantitative information of thermophysical parameters, spatial information of thermal anomaly regions, and optical information of polarization characteristics, forming a comprehensive feature representation that fully reflects the surface polishing state. To improve the efficiency of subsequent model processing, principal component analysis can be performed on the one-dimensional vector to reduce the vector dimension while retaining core feature information. The final output comprehensive feature representation combines information integrity and dimensionality economy, providing standardized input for the evaluation of surface polishing quality.

[0075] In another preferred embodiment of the present invention, the specific process of inputting the comprehensive feature representation into the hierarchical model and calculating the matching degree with the standard feature template of each polishing quality grade in step S6 is as follows:

[0076] The construction of the standard feature template library for polishing quality grades relies on standardized sample data. Standard shell samples covering different polishing quality grades are selected, and comprehensive feature representations are extracted one by one according to a predetermined feature extraction process to form the baseline feature vector for each grade. Sample selection must ensure representativeness; each grade includes multiple sets of sample features from different batches with the same process parameters. Individual differences are eliminated through mean fusion to obtain the standard feature template for that grade. The template library is stored in a categorized manner according to grade number. Each template is accompanied by corresponding polishing process parameters and a quality inspection report, providing a traceability basis for subsequent matching results. Simultaneously, feature dimensions are standardized for all templates to ensure complete consistency with the dimensions and units of the features to be measured.

[0077] The matching of the comprehensive features of the shell under test with the standard template begins with the calculation of Euclidean distance. First, the feature vectors of both are aligned and verified to ensure semantic correspondence of features in each dimension, such as a perfect match in the order of thermal and polarization feature channels. Before calculation, min-max normalization is used to eliminate magnitude differences in different dimensions of the feature vectors, focusing the distance calculation on differences in feature distribution rather than numerical magnitude. The Euclidean distance is calculated based on the differences between corresponding elements in the feature space; the smaller the distance value, the higher the fit between the feature under test and the standard template.

[0078] The conversion from Euclidean distance to matching score employs a negative correlation mapping mechanism. A monotonically decreasing function maps the distance value to the interval between 0 and 1. The parameters of this mapping function are determined through standard sample testing to ensure significant differentiation in matching scores between templates of different levels. For cases where the distance exceeds a preset threshold, the matching score is set to an extremely low value, marked as requiring review, indicating a possible abnormal polishing state.

[0079] The final evaluation result is determined by ranking the matching degree scores, and the grade corresponding to the standard template with the highest score is selected as the polishing quality grade of the shell to be tested. In extreme cases where multiple grades have the same score, auxiliary indicators such as the proportion of thermal anomaly area and the average polarization modulation depth are introduced for secondary judgment. The evaluation report is generated in a structured format, clearly indicating the polishing quality grade number, the matching degree score ranking of each standard template, and adding a key feature difference analysis between the tested feature and the optimal matching template, such as the deviation of the thermal diffusion time constant and the regional differences of polarization characteristics, to provide data support for the adjustment of the polishing process and ensure the practicality and traceability of the evaluation results.

[0080] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An image processing method for evaluating the polishing effect of an enclosure, characterized in that, Includes the following steps: S1. Acquire dynamic thermal imaging sequences of the outer shell surface under controllable thermal excitation conditions, including basic state images before thermal excitation begins, temperature change image sequences during thermal excitation, and temperature recovery images after thermal excitation stops. S2. Extract the response curve of temperature change over time for each pixel in the dynamic thermal imaging sequence, and divide the outer shell surface into a region of consistent thermal conduction and a region of thermal anomaly based on the morphological characteristics of the response curve. S3. Analyze the temperature change pattern in the heat conduction uniformity region, calculate the surface temperature rise gradient, surface temperature distribution uniformity index and heat diffusion time constant, and combine them to form a set of thermophysical parameters. S4. During the thermal excitation process, several acquisition times are set to synchronously acquire polarized scattered light images of the outer shell surface and extract polarization image features. S5. The set of thermophysical parameters, the spatial distribution information of thermal anomaly regions, and polarization image features are fused together to generate a comprehensive feature representation of the surface polishing state. S6. Input the comprehensive feature representation into the grading model, calculate the matching degree with the standard feature template of each polishing quality level, determine the polishing quality level based on the matching degree, and generate an evaluation report. In step S5, the specific process of fusing the set of thermophysical parameters, the spatial distribution information of the thermal anomaly region, and the polarization image features to generate a comprehensive feature representation of the surface polishing state is as follows: The surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant in the thermophysical parameter set are mapped to three independent thermal feature channels, and the three thermal feature channels are combined into a thermal feature cube. The spatial distribution information of the thermal anomaly region is converted into a spatial weighted distribution map, and the spatial weighted distribution map is used to perform region-selective enhancement of the polarization modulation depth distribution map and the depolarization degree distribution map. The enhanced polarization modulation depth distribution map and depolarization degree distribution map are concatenated with the thermal feature cube at the channel level to form a multimodal feature cube. A three-dimensional convolution operation is performed on the multimodal feature cube to extract the joint features of the spatial-feature dimensions. The joint features are flattened into a one-dimensional vector as a comprehensive feature representation of the surface polishing state. The specific process of using the spatial weight distribution map to perform region-selective enhancement of the polarization modulation depth distribution map and the depolarization degree distribution map is as follows: The spatial weight distribution map is multiplied by the surface temperature distribution uniformity index channel in the thermal feature cube to obtain the enhanced weight distribution map. The enhanced weight distribution map is then subjected to a nonlinear transformation. The transformed enhanced weight distribution map is then multiplied pixel by pixel with the polarization modulation depth distribution map and the depolarization degree distribution map to obtain the region-enhanced polarization modulation depth distribution map and the region-enhanced depolarization degree distribution map. The correlation coefficient matrix between the polarization modulation depth distribution map and the depolarization degree distribution map of the region enhancement is calculated. The correlation coefficient matrix is ​​then used to perform feature interaction with the thermal diffusion time constant channel in the thermal feature cube. The feature interaction process is implemented by matrix element-wise multiplication. The local correlation coefficient matrix is ​​multiplied element-wise with the thermal diffusion time constant channel in the thermal feature cube to generate a polarization feature representation with thermal-optical coupling characteristics.

2. The image processing method for evaluating the polishing effect of a housing according to claim 1, wherein In step S2, the specific process of dividing the outer shell surface into a thermally consistent region and a thermally abnormal region based on the morphological characteristics of the response curve is as follows: For each pixel location in the dynamic thermal imaging sequence, complete temperature data from the start of thermal excitation to the end of temperature recovery is extracted to form a temperature time series. Multimodal curve fitting is performed on the temperature time series to obtain a temperature change curve containing multiple components. The first derivative sequence of the temperature change curve is calculated to obtain the temperature change rate at each moment. The global maximum and global minimum points are located on the temperature change rate curve. The amplitude difference between the maximum and minimum points is calculated to determine the zero-crossing point where the temperature change rate curve changes from a positive value to a negative value. The time offset of the zero-crossing point relative to the moment when thermal excitation stops is recorded. Pixels with amplitude differences within a predetermined range and time offsets that meet predetermined standards are classified as thermal conduction consistency regions. Pixels whose amplitude difference exceeds the predetermined range or whose time offset does not meet the predetermined standard are classified as thermal anomalous pixels. Spatial connectivity analysis is performed on thermal anomalous pixels to merge adjacent thermal anomalous pixels into continuous thermal anomalous regions.

3. The image processing method for evaluating the polishing effect of a housing according to claim 2, characterized in that, The specific process of performing multimodal curve fitting on the temperature time series is as follows: A multimodal fitting equation is established that includes heat conduction, heat convection and heat radiation components. The heat conduction component adopts the form of an exponential decay function, the heat convection component adopts the form of a linear function, and the heat radiation component adopts the form of a power function. The least squares method is used to solve the coefficient parameters of each component in the multimodal fitting equation. Based on the coefficient parameters, the contribution of each component to the overall temperature response is calculated. The contribution of each component is calculated using variance contribution as an indicator. By decomposing the total variance of the temperature time series, the variance explained by each component is obtained. This proportion is the contribution of the corresponding component. The component with the largest contribution is identified as the dominant thermal response mode. Based on the type of dominant thermal response mode, the pixel regions are initially classified. Regions dominated by thermal conduction are marked as Class I regions, regions dominated by thermal convection are marked as Class II regions, and regions dominated by thermal radiation are marked as Class III regions.

4. The image processing method for evaluating the polishing effect of a housing according to claim 1, characterized in that, In step S3, the specific process of analyzing the temperature change pattern in the heat conduction uniformity region and calculating the surface temperature rise gradient, surface temperature distribution uniformity index, and heat diffusion time constant is as follows: Determine the start time of thermal excitation and the time when the temperature reaches stability, and calculate the average rate of change of surface temperature within this time interval as the surface temperature rise gradient. During the temperature stabilization phase of the thermal excitation process, the standard deviation of the temperature values ​​of all pixels within the heat conduction uniformity region is calculated as an indicator of the surface temperature distribution uniformity. During the temperature recovery phase after thermal excitation stops, the temperature drop curve is fitted with an exponential function, and the thermal diffusion time constant is extracted from the coefficients of the exponential term of the fitted function. The surface temperature rise gradient, surface temperature distribution uniformity index, and thermal diffusion time constant are normalized in terms of dimensions and combined to form a set of thermophysical parameters.

5. The image processing method for evaluating the polishing effect of a housing according to claim 1, characterized in that, In step S4, the specific process of setting several acquisition times during the thermal excitation process, synchronously acquiring polarized scattered light images from the surface of the outer shell, and extracting polarization image features is as follows: Based on the analysis of the temperature change image sequence during thermal excitation, the moments when the temperature change rate reaches its peak and when the temperature distribution pattern undergoes abrupt changes are determined. These moments are marked as acquisition moments. At each acquisition moment, a polarization light source and a polarization camera are activated for synchronous acquisition. The polarization light source provides incident light with a specific polarization direction, and the polarization camera is equipped with a rotatable analysis mirror. At each acquisition moment, a set of polarized scattered light images is acquired. The image set contains multiple scattered light images under different polarization directions from the analysis mirror. The polarization modulation depth distribution map and depolarization degree distribution map are extracted from the polarized scattered light images as polarization image features.

6. The image processing method for evaluating the polishing effect of a housing according to claim 1, characterized in that, In step S6, the specific process of inputting the comprehensive feature representation into the hierarchical model and calculating the matching degree with the standard feature template of each polishing quality level is as follows: A standard feature template library for polishing quality grades is constructed. Each template corresponds to a comprehensive feature representation of a polishing quality grade. The comprehensive feature representation of the shell to be tested is compared with each standard feature template in the template library to calculate the Euclidean distance in the feature space. The Euclidean distance is converted into a matching score. The matching score is negatively correlated with the Euclidean distance. The polishing quality grade with the highest matching score is selected as the final evaluation result, and an evaluation report is generated. The evaluation report includes the polishing quality grade number and the matching score of each standard feature template.

Images