Watch bezel surface roughness online evaluation method based on convolutional neural network

Abstract

The present application relates to the technical field of industrial machine vision detection, and specifically discloses a watch frame surface roughness online evaluation method based on a convolutional neural network, a multi-angle optical imaging system is used to collect multi-modal texture images of the watch frame surface, and an image set containing infrared reflection, visible light and polarized light information is constructed through image registration; then, frequency domain transformation analysis is performed on the images, surface micro-topography spectrum features are extracted, texture consistency eigenvalues are calculated, an optical response abnormal distribution matrix is constructed through pixel intensity response analysis, and defect sensitivity eigenvalues are calculated; the two eigenvalues are fused into a comprehensive roughness feature vector, which is input into a pre-trained convolutional neural network model for multi-feature fusion analysis; and finally, automatic determination of three levels is realized according to the output result.

Description

Technical Field

[0001] This invention relates to the field of industrial machine vision inspection technology, specifically to an online evaluation method for the surface roughness of watch bezels based on convolutional neural networks. Background Technology

[0002] With the rapid development of wearable devices such as smartwatches, consumers are increasingly demanding higher product appearance quality. As an aesthetic component, the surface roughness of a watch bezel directly affects the product's appearance and texture. Traditional surface roughness detection mainly relies on contact profilometers, which require direct contact between the probe and the surface, resulting in low detection efficiency, potential scratches on the product surface, and the inability to achieve full inspection. In recent years, non-contact inspection methods based on machine vision have gradually been applied, but existing technologies still have significant limitations: images under single illumination conditions cannot fully reflect the surface's microscopic morphology; traditional image processing algorithms have limited ability to express complex texture features; there is a lack of effective methods for multi-feature fusion, leading to insufficient detection accuracy and stability; and existing systems struggle to meet the dual requirements of detection speed and reliability on industrial production lines.

[0003] Especially in the high-end watch manufacturing industry, surface treatment processes are complex, involving both regular textures such as brushed and sandblasted finishes, as well as various minute defects. Existing visual inspection methods often fail to accurately distinguish between process textures and actual defects, easily leading to misjudgments. Furthermore, environmental changes on the production line and product positioning deviations can also affect the stability of inspection results. Therefore, there is an urgent need to develop a surface roughness assessment method that can adapt to industrial production environments and achieve high-precision online inspection. Summary of the Invention

[0004] The purpose of this invention is to provide an online evaluation method for the surface roughness of watch bezels based on convolutional neural networks, in order to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] An online evaluation method for the surface roughness of a watch bezel based on a convolutional neural network includes the following steps:

[0007] S1: Acquire images of the watch bezel surface to obtain a set of surface texture images of the watch bezel under multiple lighting angles;

[0008] S2: Perform frequency domain transformation analysis on the surface texture image set of the watch bezel, construct the surface micromorphology spectrum features, and calculate the surface texture consistency feature value;

[0009] S3: Extract the pixel intensity response at the same location in the surface texture image of the watch bezel under different lighting angles, construct the optical response anomaly distribution matrix, and calculate the surface defect sensitivity feature value;

[0010] S4: The surface texture consistency feature value and the surface defect sensitivity feature value are fused into a comprehensive roughness feature vector, which is then input into a pre-trained convolutional neural network evaluation model for multi-feature fusion analysis;

[0011] S5: Determine the surface roughness level of the watch bezel based on the output of the evaluation model: when the comprehensive roughness characteristic value is lower than the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; when it is higher than the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered.

[0012] As a further aspect of the present invention: the acquisition process of the surface texture image set specifically includes:

[0013] A high-resolution industrial area array camera is used; infrared light, visible light and multi-band polarized light illumination modes with different illumination angles are triggered sequentially by controlling the light source; the camera exposure is triggered synchronously in each illumination mode to capture the reflection characteristics of the watch bezel surface; the images acquired in different illumination modes are aligned; and finally, a surface texture image set containing multi-dimensional optical properties is generated for subsequent analysis and processing.

[0014] As a further aspect of the present invention: the frequency domain transformation analysis of the surface texture image set of the watch bezel to construct the surface micromorphological spectrum features specifically includes:

[0015] A baseline texture image of the watch bezel surface under a specific illumination angle is acquired. The baseline texture image is preprocessed, including flat field correction and noise suppression. The preprocessed image is converted from the spatial domain to the frequency domain to obtain a two-dimensional spectral distribution. The mid-frequency ring band energy distribution characterizing the periodic texture features of the surface is extracted from the two-dimensional spectrum. At the same time, the high-frequency speckle energy distribution characterizing the random surface features is extracted. The mid-frequency ring band energy and the high-frequency speckle energy are normalized and fused according to frequency band partitions to form a frequency domain feature vector. The frequency domain feature vector is mapped to a preset surface morphology feature space to generate a physically meaningful surface micromorphology spectrum feature.

[0016] As a further aspect of the present invention: the process for obtaining the surface texture consistency feature value is as follows:

[0017] The surface micromorphological spectral features of multiple detection areas on the watch bezel are selected; the cosine similarity and Euclidean distance between the morphological spectral features of each area are calculated; the similarity and distance dimensions are unified and integrated into a local consistency index; the distribution of the local consistency index of the entire surface is calculated using a sliding window; the distribution of the local consistency index is Gaussian fitted, and the kurtosis and skewness features of the fitted curve are extracted; the kurtosis and skewness are input into a pre-trained texture consistency evaluation model, and the output is quantized into surface texture consistency feature values ​​between zero and one.

[0018] As a further aspect of the present invention: the construction of the optical response anomaly distribution matrix specifically includes:

[0019] Image sequences of the same detection area on the watch bezel surface under different illumination angles were selected; intensity response curves of all pixels in the corresponding area under multi-angle illumination were established; the fitting residuals of each response curve with the ideal diffuse reflection model were calculated; the fitting residual values ​​of each pixel were arranged according to their original spatial positions; adaptive threshold segmentation was used to extract residual abnormal points; a two-dimensional abnormal distribution map was constructed based on the spatial distribution density of abnormal points; morphological opening and closing operations were performed on the two-dimensional abnormal distribution map to eliminate noise interference; finally, the processed two-dimensional abnormal distribution map was quantized into an optical response abnormal distribution matrix.

[0020] As a further aspect of the present invention: the process for obtaining the surface defect sensitivity characteristic value is as follows:

[0021] Regional connectivity analysis is performed on the optical response anomaly distribution matrix to identify potential defect regions; the area, perimeter, and anomaly intensity integral value of each potential defect region are calculated; the defect regions are classified into scratch, pit, or pockmark types according to their morphological characteristics; sensitivity weight coefficients for each type of defect are established based on historical defect data; the anomaly intensity integral value of the defect region is multiplied by the corresponding sensitivity weight coefficient to obtain the regional defect sensitivity index; all regional defect sensitivity indices are weighted, summed, and normalized to finally obtain surface defect sensitivity characteristic values ​​in the range of zero to one hundred.

[0022] As a further aspect of the present invention: the process of constructing the comprehensive roughness feature vector is as follows:

[0023] Logarithmic transformation is applied to the surface texture consistency feature values; Gaussian normalization is applied to the surface defect sensitivity feature values ​​to eliminate dimensional differences; weighted fusion of the texture consistency feature values ​​and surface defect sensitivity feature values ​​is performed based on weight coefficients obtained from historical samples; the weighted feature values ​​are mapped to a unified feature space; finally, the mapping result is combined with the original texture consistency feature values ​​and surface defect sensitivity feature values ​​to form a three-dimensional comprehensive roughness feature vector.

[0024] As a further aspect of the present invention: the construction process of the pre-trained convolutional neural network evaluation model specifically includes:

[0025] A lightweight network body is constructed using a depthwise separable convolutional structure to reduce computational complexity. The network input layer receives a three-dimensional comprehensive roughness feature vector. The first hidden layer uses a one-dimensional convolutional kernel for feature enhancement. The second hidden layer uses gated recurrent units to capture the temporal dependencies between features. The output layer uses a softmax activation function to generate a three-class classification probability output. The model training uses a focus loss function to solve the sample imbalance problem and uses an adaptive moment estimation algorithm for parameter optimization. Finally, the parameters are solidified after training with a large number of samples to obtain a pre-trained convolutional neural network evaluation model.

[0026] As a further aspect of the present invention: the output result of the multi-feature fusion analysis is:

[0027] When the comprehensive roughness characteristic value is lower than the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; when it is higher than the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered.

[0028] The beneficial effects of this invention are:

[0029] (1) By employing a multi-angle optical imaging system and frequency domain transformation analysis, this method can effectively extract the microscopic morphological spectrum features and texture consistency feature values ​​of the watch bezel surface. Simultaneously, an optical response anomaly distribution matrix is ​​constructed using pixel intensity response analysis, and surface defect sensitivity feature values ​​are calculated. This method combines multiple optical information sources, making surface roughness assessment more comprehensive and accurate. A pre-trained convolutional neural network model is used to fuse these feature values ​​and perform comprehensive analysis, enabling rapid and accurate determination of surface roughness levels. Compared to traditional detection methods, this automated and intelligent method not only improves detection speed but also significantly enhances detection accuracy.

[0030] (2) This invention can evaluate the surface roughness of watch bezels in real time on the production line and automatically trigger corresponding sorting signals based on the evaluation results, achieving immediate separation of defective products. This helps reduce the defect rate on the production line and ensures the consistency and stability of product quality. The system also establishes a comprehensive alarm mechanism and quality traceability mechanism. Once continuous defective products or abnormal equipment operation are detected, the system will immediately issue an alarm and suspend testing to ensure that the problem is dealt with in a timely manner. In addition, all detection data and judgment results will be recorded in the database, which facilitates subsequent quality statistical analysis and process improvement, thereby supporting enterprises to continuously improve product quality and production efficiency. Attached Figure Description

[0031] The invention will now be further described with reference to the accompanying drawings.

[0032] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0033] 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.

[0034] Please see Figure 1 As shown, this invention is an online evaluation method for the surface roughness of a watch bezel based on a convolutional neural network, comprising the following steps:

[0035] S1: Acquire images of the watch bezel surface to obtain a set of surface texture images of the watch bezel under multiple lighting angles;

[0036] S2: Perform frequency domain transformation analysis on the surface texture image set of the watch bezel, construct the surface micromorphology spectrum features, and calculate the surface texture consistency feature value;

[0037] S3: Extract the pixel intensity response at the same location in the surface texture image of the watch bezel under different lighting angles, construct the optical response anomaly distribution matrix, and calculate the surface defect sensitivity feature value;

[0038] S4: The surface texture consistency feature value and the surface defect sensitivity feature value are fused into a comprehensive roughness feature vector, which is then input into a pre-trained convolutional neural network evaluation model for multi-feature fusion analysis;

[0039] S5: Determine the surface roughness level of the watch bezel based on the output of the evaluation model: when the comprehensive roughness characteristic value is lower than the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; when it is higher than the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered.

[0040] In S1, images of the watch bezel surface are acquired to obtain a set of surface texture images of the watch bezel under multiple lighting angles, specifically including:

[0041] Image acquisition utilizes a high-resolution industrial area scan camera paired with a programmable multi-angle ring light source system. The industrial area scan camera employs a global shutter sensor with a resolution of at least 5 megapixels and a frame rate of at least 30 frames per second, connected to a host computer via an Ethernet interface. The programmable multi-angle ring light source system consists of multiple independently controlled LED units, each with individually adjustable emission angle, intensity, and wavelength, supporting infrared, visible light, and multi-band polarized light modes. Optical filters, including infrared transmission filters, polarization filters, and neutral density filters, are mounted in front of the camera lens. The mechanical positioning device uses a high-precision servo motor-driven watch bezel clamp to ensure stable positioning during inspection. The control system, implemented by an embedded industrial computer, coordinates the synchronized operation of light source triggering, camera exposure, and mechanical movement.

[0042] The triggering and image capture process for multiple illumination modes is achieved through precise time-series control. The control system initializes the light source and camera parameters, sets the triggering sequence and duration of different illumination modes, typically in the order of infrared light mode, visible light mode, and multi-band polarized light mode, with the illumination duration for each mode controlled between 10 and 50 milliseconds. In infrared light mode, the light source system triggers an infrared LED unit of a specific wavelength to illuminate the watch bezel surface at a low angle, and the camera simultaneously triggers exposure, with the exposure time dynamically adjusted according to the infrared light intensity. In visible light mode, the light source system switches to a white LED unit to illuminate the surface at a vertical angle, and the camera triggers exposure again. In multi-band polarized light mode, the light source system sequentially triggers linearly polarized light and circularly polarized light modes, and the camera captures the corresponding images through polarization filters. In each illumination mode, the camera acquires multiple images and synthesizes a high-quality image using a weighted average method.

[0043] Image registration and alignment employs sub-pixel registration based on feature points. First, images acquired under different lighting conditions are preprocessed, including grayscale conversion, histogram equalization, and noise filtering. Then, key feature points are extracted from the images, and an accelerated, robust feature detection algorithm is used to detect corner points, edge intersections, and high-contrast regions of surface textures. A random sampling consensus algorithm is used to calculate the transformation matrix between feature points, including translation, rotation, and scaling parameters. Least squares optimization ensures sub-pixel level registration accuracy. Finally, bilinear interpolation is used to resample the images, aligning them to the same coordinate system under different lighting conditions.

[0044] The generation of the multi-dimensional optical property image set is accomplished through further processing of the aligned image data. The image set includes infrared reflectance images, visible light color images, linearly polarized light images, and circularly polarized light images. Infrared reflectance images are used to analyze surface material uniformity and minute deformations caused by thermal effects. Visible light color images are used to evaluate surface color consistency and macroscopic texture features. Linearly polarized light images enhance the visibility of surface micro-undulations and defects by calculating polarization maps. Circularly polarized light images are used to eliminate specular reflections and extract true surface texture information. The image set is stored as a multi-dimensional array, with each pixel containing intensity values, polarization parameters, and texture descriptors for multiple bands, providing a data foundation for subsequent surface roughness feature extraction.

[0045] In S2, frequency domain transform analysis is performed on the surface texture image set of the watch bezel to construct the surface micromorphology spectrum features and calculate the surface texture consistency feature value, specifically including:

[0046] First, a baseline texture image is acquired and preprocessed. An image of the watch bezel surface captured under a specific lighting angle is selected as the baseline texture image; this angle must clearly reveal the main texture features of the surface. The acquired baseline texture image undergoes flat-field correction processing. By capturing a uniform standard white board image, the correction coefficient for each pixel is calculated to eliminate the effects of uneven lighting and lens vignetting. A nonlinear filtering algorithm is then used to suppress noise in the image, reducing random noise interference while preserving texture details. The preprocessed image retains the true surface texture information, providing a high-quality data foundation for subsequent frequency domain analysis.

[0047] Next, a spatial-domain to frequency-domain transformation is performed. A two-dimensional discrete Fourier transform is applied to the preprocessed baseline texture image to convert the image from the spatial domain to the frequency domain, obtaining the corresponding two-dimensional spectral distribution. This spectral distribution reflects the intensity information of different spatial frequency components in the image, where low-frequency components correspond to smooth regions of the image, mid-frequency components correspond to periodic texture features, and high-frequency components correspond to details and edge information. By centering the spectrum, the zero-frequency component is moved to the center of the spectrum, facilitating subsequent frequency band feature analysis.

[0048] The frequency domain feature extraction process includes the calculation of mid-frequency ring band energy and high-frequency speckle energy. In the two-dimensional spectrum, a ring region with a specific radius is defined as the mid-frequency ring band, with the center of the spectrum as the origin. The sum of the energies of all frequency components within this region is calculated; this energy value reflects the intensity characteristics of the surface's periodic texture. Simultaneously, the high-frequency speckle energy is calculated in the peripheral region of the spectrum; this energy value characterizes the distribution of the surface's random texture features. By adjusting the radius and width of the ring band, adaptive analysis can be performed on the surface texture features of watch bezels of different sizes.

[0049] The construction of the frequency domain feature vector is accomplished through the following steps: the mid-frequency ring band energy and the high-frequency speckle energy are normalized to eliminate the difference in dimensions; according to the frequency band partitioning results, the energy values ​​of different frequency bands are fused according to specific weights; finally, a frequency domain feature vector containing multiple dimensions is formed, which comprehensively characterizes the distribution characteristics of surface texture in the frequency domain.

[0050] The generation of surface micromorphology spectral features is achieved through a mapping process. A surface morphology feature space trained on a large number of samples is established, which contains the frequency domain feature distribution patterns of various typical surface morphologies. The frequency domain feature vector of the image to be tested is projected into this feature space, and by calculating the distance and similarity with various typical features, surface micromorphology spectral features with clear physical meaning are generated. This feature is represented in the form of a multi-dimensional vector, with each dimension corresponding to a surface morphology characteristic.

[0051] The calculation of surface texture consistency feature values ​​begins with the selection of multiple detection regions. Several detection regions are uniformly selected on the surface of the watch bezel, each containing complete surface texture information. These regions are ensured to be spatially representative and reflect the overall texture feature distribution of the surface. For each selected detection region, its surface micromorphological spectrum features are extracted to form a feature set.

[0052] Similarity metrics include the calculation of cosine similarity and Euclidean distance. For any two detected regions, the cosine similarity between them is calculated, reflecting the directional consistency of the feature vectors. Simultaneously, the Euclidean distance is calculated, reflecting the degree of absolute difference between the feature vectors. By combining these similarity metrics, the consistency of surface texture between different regions can be comprehensively evaluated.

[0053] The fusion of local consistency indices is achieved through dimensional unification and weighted summation. Cosine similarity is converted to a distance dimension, making it comparable to Euclidean distance. An adaptive weight allocation strategy is employed to adjust the contribution of different metrics based on surface texture characteristics. Finally, a comprehensive local consistency index is generated; a higher index value indicates greater texture consistency between two regions.

[0054] Sliding window analysis is used to evaluate the uniformity distribution across the entire surface. A sliding window is defined on the watch bezel surface, with its size adaptively determined based on the surface dimensions and texture features. The window is slid sequentially to each position, and the mean and variance of the local uniformity index for all detected areas within the window are calculated. By recording the statistics at each window position, a distribution map of the local uniformity index across the entire surface is obtained.

[0055] Distribution feature extraction is achieved by Gaussian fitting of the local consistency index distribution. The least squares method is used to fit the actual distribution to a Gaussian distribution curve, and the goodness of fit is calculated to evaluate the fitting quality. Kurtosis, reflecting the sharpness of the distribution curve, is extracted from the fitted curve; skewness, reflecting the degree of asymmetry in the distribution, is also extracted. These feature parameters can effectively characterize the overall distribution characteristics of surface texture consistency.

[0056] The final feature value quantization is performed using a pre-trained evaluation model. This model is trained using a machine learning algorithm, taking kurtosis and skewness features as input and outputting surface texture consistency feature values ​​between zero and one. Feature values ​​closer to one indicate higher surface texture consistency, while values ​​closer to zero indicate lower consistency. The model parameters are optimized through extensive training on numerous samples, accurately reflecting the non-linear mapping relationship between surface texture consistency and feature parameters.

[0057] In S3, the pixel intensity response at the same location in the surface texture image of the watch bezel under different lighting angles is extracted, an optical response anomaly distribution matrix is ​​constructed, and surface defect sensitivity feature values ​​are calculated, specifically including:

[0058] First, multi-angle image sequences were acquired and registered. Image sequences of the same detection area on the watch bezel surface were acquired under eight different lighting angles, ranging from 0 to 70 degrees, and evenly distributed at 10-degree intervals. Three images were acquired for each lighting angle, and a high-quality image was synthesized using a weighted average method to eliminate the influence of random noise. An image registration algorithm based on SIFT feature points was used to precisely align all images to the same coordinate system, achieving sub-pixel accuracy and ensuring that each pixel corresponds to the same physical location in different images.

[0059] Next, pixel intensity response curves are constructed. For the registered image sequence, the intensity value of each pixel at eight illumination angles is extracted to form an intensity response curve. The horizontal axis of the curve represents the illumination angle, and the vertical axis represents the normalized pixel intensity value. The intensity values ​​are represented by 16-bit integers, ranging from 0 to 65535. Each response curve is smoothed using a Savitzky-Golay filter with a window size of 5 and a polynomial order of 3, eliminating measurement noise while preserving curve characteristics.

[0060] The ideal diffuse reflection model is established based on Lambert's law. This model assumes the surface is an ideal diffuse reflector, and the theoretical intensity response at different illumination angles is proportional to the cosine of the incident angle. For each pixel, the theoretical intensity value is calculated based on the normal direction of its surface, generating an ideal response curve. The theoretical curve and the actual measured curve use the same illumination angle sequence and normalization method to ensure comparability.

[0061] The least squares method was used to calculate the fitting residuals. The actual intensity response curve of each pixel was fitted to the ideal diffuse reflection model curve, and the residual value was calculated for each illumination angle. The residual value is the absolute difference between the actual intensity value and the theoretical intensity value, stored as a 32-bit floating-point number. The sum of squares of the residuals for all illumination angles was used to obtain the total fitting residual for that pixel. The goodness of fit was evaluated using the coefficient of determination; the closer the value is to 1, the better the fit.

[0062] Residual anomaly detection employs an adaptive thresholding method. First, the mean and standard deviation of the residual values ​​for all pixels are calculated, and a threshold is set to the mean plus three times the standard deviation. Pixels exceeding this threshold are marked as anomalies. To adapt to the characteristics of different regions, the image is divided into 32×32 pixel blocks, and a threshold is calculated separately for each block, improving the adaptability of the detection.

[0063] The construction of a two-dimensional anomaly distribution map is based on the spatial distribution of anomaly points. The residual values ​​of each anomaly point are arranged according to their original image coordinates to generate a two-dimensional distribution map of the same size as the original image. Missing positions are filled with zero values ​​to form a complete distribution map. The distribution map is then Gaussian smoothed with a standard deviation set to 1.5 pixels to reduce the impact of isolated noise points.

[0064] Morphological processing employs an opening-then-closing operation order. The opening operation uses a circular structuring element with a radius of 3 pixels to eliminate small, isolated outliers. The closing operation uses the same structuring element to fill small holes in the outlier region. After morphological processing, the outlier distribution area is more complete and coherent, and noise is effectively suppressed.

[0065] The optical response anomaly distribution matrix is ​​generated by quantizing the processed distribution map. The anomaly intensity value of each pixel is quantized into an 8-bit integer, ranging from 0 to 255. A larger value indicates a higher degree of anomaly at that location. Finally, a grayscale image matrix with the same size as the original image is generated as the optical response anomaly distribution matrix.

[0066] The region connectivity analysis employs the 8-neighborhood connection criterion. The optical response anomaly distribution matrix is ​​scanned, and adjacent anomalous pixels are merged into connected regions. Each connected region represents a potential surface defect area. The starting coordinates, area, and outer contour information of each region are recorded. Area is calculated using pixel counting, and perimeter is calculated using chain code.

[0067] Defect feature extraction includes geometric features and intensity features. Geometric features calculate the area, perimeter, circularity, and aspect ratio of each connected region. Intensity features calculate the sum, average, and maximum of all pixel outliers within the region. These features are used for subsequent defect classification and severity assessment.

[0068] Defect classification is based on morphological features and intensity distribution. Scratch defects typically have a large aspect ratio and low roundness, pit defects appear as approximately circular areas with uniform intensity distribution, and pockmark defects are small-area dotted distributions with high intensity values. Using a decision tree algorithm for classification, the accuracy rate reaches over 95%.

[0069] The sensitivity weighting coefficients are derived from historical defect data. 1000 confirmed defect samples were collected, and a weighting mapping table was established based on defect type and severity. The weighting coefficient for scratches is 0.8, for dents it is 1.2, and for pits it is 1.0. The weighting coefficients are updated quarterly to maintain consistency with actual conditions.

[0070] The regional defect sensitivity index is calculated using a weighted integral method. The integral value of the anomaly intensity of each defect region is equal to the sum of the anomaly values ​​of all pixels within that region. This integral value is multiplied by the weighting coefficient of the corresponding defect type to obtain the regional defect sensitivity index. The index value is represented as a 32-bit floating-point number; a larger value indicates a greater impact of the defect on surface quality.

[0071] The final eigenvalue is calculated by weighted summation of all regional indices. The summation weights are determined based on the proportion of the defect area to the total abnormal area. The summation result is then linearly mapped to a range of 0 to 100 to obtain the surface defect sensitivity eigenvalue.

[0072] In S4, surface texture consistency feature values ​​and surface defect sensitivity feature values ​​are fused into a comprehensive roughness feature vector, which is then input into a pre-trained convolutional neural network evaluation model for multi-feature fusion analysis, specifically including:

[0073] The construction of the comprehensive roughness feature vector begins with feature preprocessing. The surface texture consistency feature values ​​are first transformed using a natural logarithmic transformation. This uses a logarithmic function with the natural constant e as the base to convert the feature value distribution into an approximately normal distribution, improving data stability. The surface defect sensitivity feature values ​​are Gaussian normalized. The mean and standard deviation of these feature values ​​in the training sample set are calculated, and then standardized by subtracting the mean and dividing by the standard deviation to eliminate dimensional differences and ensure the values ​​are distributed within a range with a mean of 0 and a standard deviation of 1.

[0074] Feature-weighted fusion employs weighting coefficients learned from historical samples. 1000 labeled samples were collected, and through correlation analysis and feature importance assessment, the weighting coefficient for surface texture consistency was determined to be 0.6, and the weighting coefficient for surface defect sensitivity was determined to be 0.4. The weighted fusion formula is: the fused feature value equals the texture consistency feature value multiplied by 0.6 plus the defect sensitivity feature value multiplied by 0.4. The weighting coefficients are recalculated and updated every six months based on newly collected sample data.

[0075] The feature space mapping employs a nonlinear transformation method. The weighted and fused feature values ​​are input into a hyperbolic tangent activation function, mapping the feature values ​​to the range of -1 to 1. Simultaneously, the original texture uniformity feature values ​​and defect sensitivity feature values ​​are preserved; these three values ​​together constitute a three-dimensional feature vector. This vector comprehensively represents the overall surface roughness, containing both texture uniformity information and defect severity information.

[0076] The convolutional neural network evaluation model is constructed using a depthwise separable convolutional structure. The network input layer is designed to receive a 3D comprehensive roughness feature vector with an input dimension of 1×3. The first hidden layer uses 32 one-dimensional convolutional kernels for feature enhancement, with a kernel size of 3, a stride of 1, and the ReLU activation function. This layer extracts local patterns and relationships of features through convolutional operations.

[0077] The second hidden layer employs a gated recurrent unit (GRU) structure with 16 units. This layer is used to capture potential temporal dependencies between features. Although the input features themselves are not temporally sequential, the GRU can learn long-term dependencies between features, improving the model's expressive power. The output of this layer is passed to the output layer through a fully connected layer.

[0078] The output layer uses the softmax activation function to generate three-class probability outputs. The three output nodes correspond to the three categories: superior, acceptable, and unacceptable. The output value of each node represents the probability that the input sample belongs to the corresponding category, and the sum of the three output values ​​is 1. The model ultimately selects the category with the highest probability value as the prediction result.

[0079] The model training employs a focus loss function to address the imbalanced sample problem. This focus loss function introduces a moderating factor on top of the standard cross-entropy loss, reducing the weights of easily classified samples and allowing the model to focus more on difficult-to-classify samples. The moderating factor is set to 2, and the balancing parameter is set to 0.25. Parameter optimization is performed using an adaptive moment estimation algorithm, with an initial learning rate of 0.001, a beta1 parameter of 0.9, a beta2 parameter of 0.999, and an epsilon parameter of 1e-8.

[0080] The model training process employs a 5-fold cross-validation method. The training dataset is randomly divided into 5 subsets of equal size. Four subsets are used for training and one subset for validation in each iteration, repeated 5 times. The training period is set to 100 epochs, and the batch size is set to 32. After each epoch, the accuracy on the validation set is calculated. Training is terminated early if the validation set accuracy does not improve for 10 consecutive epochs.

[0081] After training, the model parameters that perform best on the validation set are selected and solidified. These parameters include all trainable parameters such as convolutional kernel weights, recurrent unit weights, fully connected layer weights, and bias terms. The solidified model is saved as a binary file and deployed to an embedded device for online evaluation.

[0082] After model deployment, a regular update mechanism is established. Every 500 new labeled samples collected, these samples are used to fine-tune the model and update its parameters. The learning rate for fine-tuning is set to one-tenth of the initial learning rate, and the training cycle is shortened to 20 epochs. This ensures the model can adapt to changes in data distribution caused by variations in production processes.

[0083] During the online evaluation, the model receives the comprehensive roughness feature vector as input, completes the forward propagation calculation within 10 milliseconds, and outputs probability values ​​for three categories. The system is equipped with a dual threshold judgment mechanism: when the probability of a superior product is higher than 0.7, it is directly judged as a superior product; when the probability of a defective product is higher than 0.6, it is directly judged as a defective product; when all probability values ​​are lower than the threshold, a review mechanism is activated, requiring data to be collected again for secondary analysis.

[0084] The final output includes a roughness grade code, a quality judgment conclusion, and a test confidence index. The roughness grade code uses 0, 1, and 2 to represent superior, qualified, and unqualified products, respectively. The quality judgment conclusion is a descriptive string. The test confidence index is the maximum probability value output by the model, indicating the reliability of the test. All results are transmitted to the production line control system in real time via an industrial bus.

[0085] The system establishes a comprehensive quality traceability mechanism. Information such as the original feature values, model input and output data, final judgment results, and timestamps for each test are recorded in the database. This data is used for quality statistical analysis, model performance monitoring, and process improvement analysis. Data is retained for two years and supports multi-dimensional querying and analysis by time range, product batch, defect type, and other dimensions.

[0086] Model performance is monitored using a continuous evaluation method. Each week, 100 samples are randomly selected from the production data for manual review. The review results are compared with the model's prediction results, and metrics such as accuracy, recall, and precision are calculated. If any metric drops by more than 5% for two consecutive weeks, the model retraining process is triggered to ensure the model consistently maintains optimal performance.

[0087] In S5, the surface roughness level of the watch bezel is determined based on the output of the evaluation model: when the comprehensive roughness characteristic value is below the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; and when it is above the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered, specifically including:

[0088] Surface roughness grade determination is based on the comprehensive roughness characteristic value output by the evaluation model. This characteristic value is a continuous value ranging from 0 to 100, with lower values ​​indicating better surface quality. The first threshold is set at 25, determined through statistical characteristic value analysis of 500 superior samples, ensuring that 95% of the superior samples have characteristic values ​​below this threshold. The second threshold is set at 60, determined through statistical analysis of 300 non-conforming samples, ensuring that 90% of the non-conforming samples have characteristic values ​​above this threshold. After the thresholds are determined, they are verified and adjusted quarterly based on newly collected sample data.

[0089] When the overall roughness characteristic value is below 25, the watch bezel is judged to be of superior quality. The system records this judgment result and marks the product as the highest quality level in the database. Simultaneously, a superior quality report is generated, including information such as the inspection time, product number, and characteristic value data. This report is automatically uploaded to the enterprise's quality management system for quality traceability and statistical analysis.

[0090] When the overall roughness characteristic value is between 25 and 60, the watch bezel is judged to be a qualified product. The system records the judgment result and marks it as a qualified quality level in the database. A qualified product quality report is generated, including detailed test data and quality scores. For critically qualified products with characteristic values ​​close to 60, the system adds an additional warning flag to prompt quality personnel to pay attention to the quality trend of such products.

[0091] When the overall roughness characteristic value exceeds 60, the watch bezel is determined to be a defective product. The system immediately records the determination result and marks it as a non-conforming quality level in the database. Simultaneously, a non-conforming product report is generated, detailing the defect type, severity, and location information. The system automatically triggers a sorting control signal, which is transmitted to the production line sorting device via industrial Ethernet.

[0092] The sorting control signal uses a 24V DC level signal with a duration of 500 milliseconds. Upon receiving the signal, the sorting device activates the rejection mechanism when a product arrives at the sorting station. This rejection mechanism employs a pneumatic pusher with a stroke of 50 mm and an action time of 100 milliseconds, ensuring accurate placement of defective products into the waste collection bin. The system simultaneously records the execution time and results of the sorting action, creating a complete quality processing record.

[0093] All judgment results and sorting operations are displayed in real time on the human-machine interface. Operators can view test result statistics, quality trend charts, and equipment operating status in real time. The system provides historical data query functions, supporting queries based on multiple conditions such as time range, product batch, and quality grade. All data is retained for 3 years and can be exported to standard format reports.

[0094] The system has established a comprehensive alarm mechanism. When five consecutive non-conforming products are detected, the system issues an audible and visual alarm to alert quality personnel for intervention. If equipment malfunctions or communication is interrupted, the system automatically pauses testing and sends fault information to maintenance personnel. A weekly quality statistical analysis report is generated, including first-pass yield, quality trend predictions, and improvement suggestions.

[0095] The working principle of this invention is as follows: A multi-modal texture image of the watch bezel surface is acquired using a multi-angle optical imaging system. After image registration, an image set containing information on infrared reflection, visible light, and polarized light is constructed. Subsequently, frequency domain transformation analysis is performed on the images to extract surface micro-morphology spectrum features and calculate texture consistency feature values. Simultaneously, an optical response anomaly distribution matrix is ​​constructed through pixel intensity response analysis, and defect sensitivity feature values ​​are calculated. These two feature values ​​are fused into a comprehensive roughness feature vector, which is then input into a pre-trained convolutional neural network model for multi-feature fusion analysis. Finally, based on the output results, three levels of automatic judgment are achieved: feature values ​​below 25 indicate superior quality, 25-60 indicate qualified quality, and values ​​above 60 indicate unqualified quality, triggering a sorting signal and completing online quality assessment and sorting control.

[0096] 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 online evaluation method for the surface roughness of a watch bezel based on a convolutional neural network, characterized in that, Includes the following steps: S1: Acquire images of the watch bezel surface to obtain a set of surface texture images of the watch bezel under multiple lighting angles; S2: Perform frequency domain transform analysis on the surface texture image set of the watch bezel to construct the surface micromorphology spectrum features, and calculate the surface texture consistency feature value. The process of obtaining the surface texture consistency feature value is as follows: Surface micromorphology spectrum characteristics of multiple detection areas on the watch bezel were selected; Calculate the cosine similarity and Euclidean distance between the morphological spectrum features of each region; unify the similarity and distance dimensions into a local consistency index; use a sliding window to calculate the local consistency index distribution of the entire surface; perform Gaussian fitting on the local consistency index distribution and extract the kurtosis and skewness features of the fitted curve; input the kurtosis and skewness into a pre-trained texture consistency evaluation model and output a surface texture consistency feature value quantized between zero and one. S3: Extract the pixel intensity response at the same location in the surface texture image of the watch bezel under different lighting angles, construct an optical response anomaly distribution matrix, and calculate the surface defect sensitivity feature value. The process of obtaining the surface defect sensitivity feature value is as follows: Regional connectivity analysis was performed on the optical response anomaly distribution matrix to identify potential defect regions; Calculate the area, perimeter, and anomaly intensity integral value for each potential defect region; Based on the morphological characteristics of the defect area, it is classified into scratches, pits or pockmarks; based on historical defect data, sensitivity weight coefficients for each type of defect are established; the abnormal intensity integral value of the defect area is multiplied by the sensitivity weight coefficient of the corresponding type to obtain the regional defect sensitivity index; the sensitivity indexes of all regions are weighted and summed and normalized to finally obtain the surface defect sensitivity feature value in the range of zero to one hundred. S4: The surface texture consistency feature value and the surface defect sensitivity feature value are fused into a comprehensive roughness feature vector, which is then input into a pre-trained convolutional neural network evaluation model for multi-feature fusion analysis; S5: Determine the surface roughness level of the watch bezel based on the output of the evaluation model: when the comprehensive roughness characteristic value is lower than the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; when it is higher than the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered.

2. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The acquisition process of the surface texture image set specifically includes: A high-resolution industrial area array camera is used; infrared light, visible light and multi-band polarized light illumination modes with different illumination angles are triggered sequentially by controlling the light source; the camera exposure is triggered synchronously in each illumination mode to capture the reflection characteristics of the watch bezel surface; the images acquired in different illumination modes are aligned; and finally, a surface texture image set containing multi-dimensional optical properties is generated for subsequent analysis and processing.

3. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The frequency domain transform analysis of the surface texture image set of the watch bezel to construct the surface micromorphology spectrum features specifically includes: A baseline texture image of the watch bezel surface under a specific illumination angle is acquired. The baseline texture image is preprocessed, including flat field correction and noise suppression. The preprocessed image is converted from the spatial domain to the frequency domain to obtain a two-dimensional spectral distribution. The mid-frequency ring band energy distribution characterizing the periodic texture features of the surface is extracted from the two-dimensional spectrum. At the same time, the high-frequency speckle energy distribution characterizing the random surface features is extracted. The mid-frequency ring band energy and the high-frequency speckle energy are normalized and fused according to frequency band partitions to form a frequency domain feature vector. The frequency domain feature vector is mapped to a preset surface morphology feature space to generate a physically meaningful surface micromorphology spectrum feature.

4. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The construction of the optical response anomaly distribution matrix specifically includes: Image sequences of the same detection area on the watch bezel surface under different illumination angles were selected; intensity response curves of all pixels in the corresponding area under multi-angle illumination were established; the fitting residuals of each response curve with the ideal diffuse reflection model were calculated; the fitting residual values ​​of each pixel were arranged according to their original spatial positions; adaptive threshold segmentation was used to extract residual abnormal points; a two-dimensional abnormal distribution map was constructed based on the spatial distribution density of abnormal points; morphological opening and closing operations were performed on the two-dimensional abnormal distribution map to eliminate noise interference; finally, the processed two-dimensional abnormal distribution map was quantized into an optical response abnormal distribution matrix.

5. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The process of constructing the comprehensive roughness feature vector is as follows: Logarithmic transformation is applied to the surface texture consistency feature values. The surface defect sensitivity feature values ​​are Gaussian normalized to eliminate dimensional differences; the texture consistency feature values ​​and surface defect sensitivity feature values ​​are weighted and fused based on the weight coefficients obtained from learning from historical samples; the weighted feature values ​​are mapped to a unified feature space; finally, the mapping result is combined with the original texture consistency feature values ​​and surface defect sensitivity feature values ​​to form a three-dimensional comprehensive roughness feature vector.

6. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The construction process of the pre-trained convolutional neural network evaluation model specifically includes: A lightweight network body is constructed using a depthwise separable convolutional structure to reduce computational complexity. The network input layer receives a three-dimensional comprehensive roughness feature vector. The first hidden layer uses a one-dimensional convolutional kernel for feature enhancement. The second hidden layer uses gated recurrent units to capture the temporal dependencies between features. The output layer uses a softmax activation function to generate a three-class classification probability output. The model training uses a focus loss function to solve the sample imbalance problem and uses an adaptive moment estimation algorithm for parameter optimization. Finally, the parameters are solidified after training with a large number of samples to obtain a pre-trained convolutional neural network evaluation model.

7. The online evaluation method for watch bezel surface roughness based on convolutional neural networks according to claim 1, characterized in that, The output result of the multi-feature fusion analysis is: When the comprehensive roughness characteristic value is lower than the first threshold, it is judged as a superior product; when it is between the first and second thresholds, it is judged as a qualified product; when it is higher than the second threshold, it is judged as a non-qualified product and a sorting control signal is triggered.

Images