Product appearance consistency detection and quality traceability method based on deep learning

Abstract

The application discloses a product appearance consistency detection and quality tracing method based on deep learning, and particularly relates to the technical field of product appearance detection and quality tracing. A standard appearance image set corresponding to a product to be detected is acquired, and the standard appearance image set is input into an appearance feature extraction network to extract standard color features, texture features, gloss features and boundary transition features according to preset appearance regions, and a standard appearance feature template is established. A to-be-detected image set of the product to be detected under the same light source direction and shooting angle is collected, and pose correction and region registration are performed according to the standard appearance feature template to obtain a target region image set. A consistency offset is calculated based on a target appearance feature template and the standard appearance feature template, an abnormal appearance region and an offset feature are determined, and an abnormal appearance feature group is formed. Then, the abnormal contribution degrees of each candidate process node are generated in combination with production process data, and a quality tracing result is output. The application improves the appearance consistency detection precision and the abnormal source positioning efficiency.

Description

Technical Field

[0001] This invention relates to the field of product appearance inspection and quality traceability technology, specifically to a deep learning-based method for product appearance consistency inspection and quality traceability. Background Technology

[0002] Product appearance consistency testing is widely used in the factory inspection of consumer electronics, automotive interiors, home appliance panels, and precision structural components. Taking the frame of a smart wearable product with a gradient color coating and narrow high-gloss chamfered edges as an example, during the sandblasting, anodizing, dyeing, sealing, cleaning, and drying processes, such products are prone to slight color differences, texture fogging, sudden changes in gloss, or edge whitening in the same batch of products due to factors such as equipment cavity temperature fluctuations, fixture slot obstruction, local concentration changes of dyeing solution, or uneven drying airflow. These types of anomalies typically do not manifest as obvious scratches, dents, or dirt, and only appear under specific light source directions and shooting angles. They are easily missed when using manual visual inspection or fixed threshold image detection. Although existing deep learning detection methods can identify and classify defects, most of them only output whether they are qualified or not or the defect category. They fail to associate the color, texture, gloss, and boundary transition offset features of the abnormal appearance area with specific process numbers, equipment numbers, fixture numbers, slot numbers, and process parameters. As a result, when local consistency offsets occur in batch products, it is difficult to quickly determine the source of the anomaly. It is necessary to rely on line stoppage for investigation and manual experience for tracing, which has a long investigation cycle, low positioning accuracy, and can easily lead to batch rework or scrap.

[0003] Therefore, how to accurately detect product appearance consistency deviations in complex appearance areas and realize quality traceability of abnormal process nodes based on the detection results has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a product appearance consistency detection and quality traceability method based on deep learning, so as to solve the shortcomings of the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a product appearance consistency detection and quality traceability method based on deep learning, comprising: Obtain a standard appearance image set and production process data corresponding to the product to be tested. The production process data includes process number, equipment number, fixture number, slot number and process parameters. The standard appearance image set is input into the appearance feature extraction network, and standard color features, texture features, gloss features and boundary transition features are extracted according to the preset appearance area to establish a standard appearance feature template. Collect a set of images of the product to be tested under the same light source direction and shooting angle, and perform posture correction and region registration on the set of images to be tested according to the standard appearance feature template to obtain the target area image set; The target region image set is input into the appearance feature extraction network to obtain the target appearance feature template. Based on the standard appearance feature template and the target appearance feature template, the consistency offset corresponding to each appearance region is calculated. When the consistency offset exceeds the threshold of the corresponding region, the abnormal appearance region and its offset feature are determined to form an abnormal appearance feature group. Based on the abnormal appearance feature group and production process data, the abnormal contribution degree of each candidate process node is generated, and the candidate process node with the largest abnormal contribution degree and its corresponding equipment number, fixture number, slot number and process parameters are used as the quality traceability result.

[0006] Preferably, the standard appearance image set and production process data corresponding to the product to be tested are obtained, including: Before the product to be inspected enters the appearance consistency inspection station, a standard appearance image set corresponding to a standard product of the same product model, color specification and appearance processing technology as the product to be inspected is obtained; the unique product identifier of the product to be inspected is read, and the corresponding process number, equipment number, fixture number, slot number and process parameters are obtained based on the unique product identifier; the standard appearance image set and production process data are bound according to the unique product identifier to form the basic inspection data corresponding to the product to be inspected.

[0007] Preferably, the acquisition of a standard appearance image set includes: Place the standard product in the positioning fixture and align the reference edge, assembly hole or outer contour of the standard product with the detection coordinate system; acquire images of the standard product according to the preset shooting angle, and light sources in different directions are lit in sequence to obtain a set of standard appearance images covering the planar area, curved area, highlight edge area and boundary transition area; record the product model, color specification, light source number, light source angle, camera number, shooting angle, exposure time and gain value for each standard appearance image.

[0008] Preferably, a standard appearance feature template is established, including: The standard appearance image set is processed by image denoising, white balance correction, distortion correction, brightness normalization, and image size unification. Based on the product structural characteristics, the appearance surface is divided into planar areas, curved areas, highlight areas, hole perimeter areas, boundary transition areas, and occlusion edge areas, and a region mask is created for each preset appearance area. The standard color features, texture features, gloss features, and boundary transition features are regionalized and summarized according to the region mask to obtain the standard appearance feature template.

[0009] Preferably, the extraction of standard color features, texture features, gloss features, and boundary transition features includes: The color feature extraction branch of the appearance feature extraction network extracts the color distribution features and color change curves within the preset appearance area; the texture feature extraction branch of the appearance feature extraction network extracts the sandblasting particle distribution features, local texture direction features, and deep texture embedding features within the preset appearance area; the gloss feature extraction branch and boundary transition feature extraction branch of the appearance feature extraction network extract the gloss response features within the preset appearance area and the boundary change features between adjacent appearance areas, respectively.

[0010] Preferably, pose correction and region registration are performed on the image set to be tested based on a standard appearance feature template, including: Call the standard boundary transition feature in the standard appearance feature template, and select the preset appearance region whose boundary transition position dispersion meets the preset sorting conditions as the stable boundary positioning region. Call the standard gloss feature in the standard appearance feature template, and select the preset appearance area whose brightness response difference under different light source directions meets the preset sorting conditions as the gloss-sensitive positioning area. The planar pose correction of the image set to be tested is performed using the boundary stable positioning region, and the arc surface tilt offset correction of the product to be tested is performed using the gloss-sensitive positioning region.

[0011] Preferably, planar pose correction of the image set under test is performed using a boundary-stabilized localization region, including: Within the stable positioning area of ​​the boundary, sampling lines are set along the direction of the standard boundary normal, and the transition position of the boundary to be tested corresponding to each sampling line is extracted from the set of images to be tested. Establish a correspondence between the transition position of the boundary to be tested and the standard boundary transition position in the standard appearance feature template to obtain the rotation offset and translation offset of the image set to be tested relative to the standard appearance image set. The image set under test is mapped in reverse coordinates based on rotation offset and translation offset to obtain the image set under test after preliminary attitude correction.

[0012] Preferably, the method of correcting the curved surface tilt offset of the product under test using a gloss-sensitive positioning area includes: The location of the specular response to be tested within the gloss-sensitive localization area is extracted from the image set after preliminary attitude correction. The specular response position to be tested is matched with the standard specular response position in the standard appearance feature template to obtain the specular residual displacement corresponding to the gloss-sensitive positioning area. Based on the specular residual displacement, coordinate compensation is performed on the preliminary attitude-corrected image set, and the image set of the target region is obtained by cropping, rotating and aligning, and unifying the size according to the region mask in the standard appearance feature template.

[0013] Preferably, the group forming the abnormal appearance feature includes: The target region image set is input into the appearance feature extraction network according to the preset appearance region order in the standard appearance feature template to extract target color features, target texture features, target gloss features and target boundary transition features that are consistent with the dimensions of the standard appearance feature template. The target appearance feature template is compared with the standard appearance feature template in the same area to obtain the color difference features, texture difference features, gloss difference features and boundary transition difference features corresponding to each preset appearance area. After subtracting the acquisition drift features common to the entire set of images under test from the color difference features and gloss difference features, the consistency offset is obtained by fusing them, and the abnormal appearance regions and their offset features are determined based on the consistency offset.

[0014] Preferably, the anomaly contribution of each candidate process node is generated based on the abnormal appearance feature group and production process data, including: Candidate process nodes are selected from the production process data according to the offset feature category, and valid parameter fragments are extracted when the product to be tested actually passes through each candidate process node. The region code of the abnormal appearance area in the product coordinates is matched with the function orientation code of the equipment number, fixture number and slot number corresponding to the candidate process node to obtain the region response matching result; Based on the regional response matching results, the deviation status of the effective parameter segments, and the direction of change of the offset characteristics, the contribution results corresponding to the equipment number, fixture number, and slot number under the same candidate process node are merged to obtain the abnormal contribution degree of each candidate process node.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention differs from existing solutions that rely solely on single-image threshold judgment or defect classification models. Its core lies in splitting a standard appearance image set into preset appearance regions and establishing regionalized standard appearance feature templates for color, texture, gloss, and boundary transitions. Instead of directly judging defects in the image under test, it first performs planar posture correction and curved surface tilt offset correction based on standard boundary transition features and standard gloss features, and then compares the images within the same region. This avoids misidentifying slight product placement deviations, curved surface reflection offsets, and differences in light source response as appearance anomalies, improving the accuracy of consistency detection for gradient coatings, high-gloss chamfers, and boundary transition areas.

[0016] This invention differs from existing appearance inspection methods that only output pass / fail or defect categories. Its core lies in combining abnormal appearance areas, offset features, and consistency offsets with production process data. By matching the effects of process numbers, equipment numbers, fixture numbers, slot numbers, and process parameters, and analyzing parameter deviations, it generates the anomaly contribution of candidate process nodes. This method allows appearance inspection results to be directly traced to specific processes, equipment, fixture slots, and abnormal process parameters, reducing downtime for item-by-item checks, shortening anomaly location time, and lowering the risk of batch rework or scrap. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the deep learning-based product appearance consistency detection and quality traceability method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] Example 1, please refer to Figure 1 As shown in this embodiment, the product appearance consistency detection and quality traceability method based on deep learning includes: Obtain a set of standard appearance images and production process data corresponding to the product to be tested. The production process data includes process number, equipment number, fixture number, slot number, and process parameters.

[0021] In some embodiments, acquiring a standard appearance image set and production process data corresponding to the product to be inspected specifically includes: before the product to be inspected enters the appearance consistency inspection station, acquiring a standard appearance image set corresponding to a standard product that belongs to the same product model, color specification, and appearance processing technology as the product to be inspected. The standard product can be a good product whose appearance consistency has been manually verified to meet preset quality requirements, or it can be a standard sample formed by averaging multiple good product images. To ensure the accuracy of subsequent appearance feature comparisons, the acquisition conditions of the standard appearance image set are consistent with the acquisition conditions of the product to be inspected.

[0022] Specifically, the standard appearance image set can be acquired using an industrial camera, telecentric lens, ring light source, bar light source, and backlight source positioned at the inspection station. During inspection, the standard product is placed on a positioning fixture, and its position is restricted by positioning pins, limiting blocks, or vacuum adsorption structures on the fixture, aligning the standard product's reference edges, assembly holes, or outer contours with the inspection coordinate system. The camera is then controlled to acquire images at a preset shooting angle, and light sources from different directions are sequentially illuminated to obtain standard appearance images of the standard product under multiple light source directions and shooting angles. For example, the standard appearance image set may include frontal illumination images, left oblique light images, right oblique light images, top grazing light images, and bottom grazing light images to cover the product's planar areas, curved areas, highlight edge areas, and boundary transition areas.

[0023] In some embodiments, each standard appearance image in the standard appearance image set carries an image acquisition identifier, which includes at least one of the following: product model, color specification, light source number, light source angle, camera number, shooting angle, exposure time, gain value, acquisition time, and image sequence number. By saving the above image acquisition identifier, the same camera parameters and light source parameters can be used when acquiring images to be tested later, thereby reducing misjudgments caused by changes in acquisition conditions.

[0024] Furthermore, the production process data corresponding to the product to be tested is obtained. This production process data can be provided by a manufacturing execution system, equipment control system, process parameter acquisition system, or barcode traceability system. In specific implementation, after the product to be tested completes each production process, its unique product identifier is read using a product QR code, laser code, RFID tag, or work order number. Based on this unique product identifier, the corresponding production process data for the product to be tested is retrieved from the production database. The production process data includes process number, equipment number, fixture number, slot number, and process parameters.

[0025] The process number indicates the specific production stage the product to be tested has undergone, such as sandblasting, anodizing, dyeing, sealing, cleaning, drying, or marking. The equipment number indicates the equipment or equipment cavity that performs the corresponding process. The fixture number indicates the jig, hanger, or tray that carries the product to be tested. The slot number indicates the specific installation position of the product to be tested on the fixture. The process parameters indicate the control parameters that affect the consistency of appearance in the corresponding process, such as at least one of sandblasting pressure, sandblasting time, oxidation voltage, oxidation current, bath temperature, dyeing time, dyeing solution concentration, sealing temperature, cleaning time, drying temperature, drying air velocity, and equipment running time.

[0026] In some embodiments, to facilitate subsequent correlation analysis between appearance anomalies and process nodes, the standard appearance image set and the production process data are bound together according to the unique product identifier of the product to be inspected, forming a basic inspection data package corresponding to the product to be inspected. The basic inspection data package includes at least the standard appearance image set, standard image acquisition parameters, product model, color specification, unique product identifier, process number, equipment number, fixture number, slot number, and process parameters. When performing subsequent steps, the inspection system directly calls the standard appearance image set in this basic inspection data package as the benchmark for appearance feature extraction and uses the production process data within it as the data source for quality traceability.

[0027] The above methods can ensure the consistency between the standard appearance image set and the product under test in terms of model, color, process and acquisition conditions. They can also ensure that the appearance inspection results of the product under test can be traced back to specific processes, equipment, fixture positions and process parameters, providing a data foundation for the subsequent establishment of standard appearance feature templates, calculation of appearance consistency offset and identification of abnormal process nodes.

[0028] The standard appearance image set is input into the appearance feature extraction network, and standard color features, texture features, gloss features and boundary transition features are extracted according to the preset appearance area to establish a standard appearance feature template.

[0029] In some embodiments, the standard appearance image set is first preprocessed. The preprocessing includes at least one of image denoising, white balance correction, distortion correction, brightness normalization, and image size unification. Specifically, standard appearance images of the same standard product under different light source directions and shooting angles can be grouped according to the light source number, shooting angle, exposure time, and camera number saved during standard appearance image acquisition, so that each group of standard appearance images corresponds to the same appearance observation conditions. Through the above preprocessing, the impact of camera response differences, local illumination variations, and lens distortion on subsequent feature extraction results can be reduced.

[0030] Preset appearance areas are determined based on the structural characteristics of the product to be inspected. These preset appearance areas can be determined using a 3D model of the product, a 2D contour template, manually annotated areas, or image segmentation results. For example, for a product with curved chamfers, planar decorative areas, highlight edge areas, and assembly boundary areas, its appearance surface can be divided into planar areas, curved areas, highlight areas, hole perimeter areas, boundary transition areas, and masking edge areas. Each preset appearance area corresponds to a region mask, which is used to define the pixel range of that area in a standard appearance image.

[0031] In some embodiments, a preprocessed set of standard appearance images is input into an appearance feature extraction network. The appearance feature extraction network can be a convolutional neural network, a visual Transformer network, or a multi-branch feature extraction network formed by combining a convolutional neural network with an attention mechanism. The appearance feature extraction network includes at least a color feature extraction branch, a texture feature extraction branch, a gloss feature extraction branch, and a boundary transition feature extraction branch. Each branch shares the shallow contour features of the standard appearance images and outputs appearance features of the corresponding type.

[0032] Specifically, the color feature extraction branch is used to extract the color distribution features of the standard appearance image within each preset appearance area. The color distribution features may include at least one of the following in the Lab color space: mean luminance, mean chromaticity, color variance, color histogram, and color gradient change. For gradient or metallic oxide products, the color features may also include a color change curve along a preset direction to characterize whether the color transition within the area is smooth.

[0033] The texture feature extraction branch is used to extract surface texture features within each preset appearance area. These texture features may include sandblasting particle distribution features, local texture direction features, texture coarseness features, gray-level co-occurrence matrix features, local binary pattern features, or deep texture embedding features output by a neural network convolutional layer. These texture features can characterize whether the product surface exhibits uneven atomization, particle accumulation, localized blooming, or inconsistent sandblasting patterns.

[0034] The gloss feature extraction branch is used to extract the reflective intensity, reflective area, specular highlight position, and highlight attenuation curve of each preset appearance area based on standard appearance images under different light source directions. In specific implementation, the image brightness response of the same preset appearance area under frontal light, side light, and low-angle grazing light conditions can be compared to obtain the gloss response characteristics of that area. These gloss response characteristics are used to characterize the gloss consistency of highlight edges, curved chamfers, or coating surfaces.

[0035] The boundary transition feature extraction branch is used to extract the boundary change features between adjacent appearance areas. The boundary transition features may include the color difference, brightness difference, texture abrupt change, edge gradient width, and transition band width on both sides of the boundary. For products with masking lines, assembly seams, marking boundaries, or high-gloss chamfers, the boundary transition features are used to characterize whether there is whitening, dark bands, staining creep, or uneven coating transition at the edge.

[0036] Furthermore, based on the region mask corresponding to each preset appearance region, the color features, texture features, gloss features, and boundary transition features are regionalized and summarized to obtain the standard region features corresponding to each preset appearance region. The standard region features can be represented as: Rj = (Mj, Cj, Pj, Gj, Bj); where Rj represents the standard region feature corresponding to the j-th preset appearance region, Mj represents the region mask of the j-th preset appearance region, Cj represents the standard color feature of the j-th preset appearance region, Pj represents the standard texture feature of the j-th preset appearance region, Gj represents the standard gloss feature of the j-th preset appearance region, Bj represents the standard boundary transition feature of the j-th preset appearance region, j = 1, 2, ..., m, and m represents the number of preset appearance regions.

[0037] In some embodiments, if the standard appearance image set includes images corresponding to multiple standard good products, then the color features, texture features, gloss features, and boundary transition features of each standard good product under the same preset appearance area are extracted, and their mean, variance, or confidence interval is calculated as the standard feature range of the preset appearance area. This avoids the influence of random deviations of a single standard sample on subsequent detection results.

[0038] A standard appearance feature template T0 is established by combining the standard region features corresponding to all preset appearance regions: T0 = {R1, R2, ..., Rj, ..., Rm}. The standard appearance feature template T0 stores the positional relationships, region masks, standard color features, standard texture features, standard gloss features, and standard boundary transition features of each preset appearance region. Subsequent posture correction, region registration, and consistency offset calculations for the product under inspection are all performed using the standard appearance feature template T0 as the comparison benchmark. Through this method, the appearance information in the standard appearance image set can be transformed into calculable and comparable regionalized standard features, providing a basis for subsequent judgment of whether the product under inspection has local appearance consistency anomalies.

[0039] The test image set of the product to be tested is collected under the same light source direction and shooting angle, and the posture correction and region registration of the test image set are performed according to the standard appearance feature template to obtain the target area image set.

[0040] In some embodiments, standard gloss features and standard boundary transition features corresponding to each preset appearance region are read from a standard appearance feature template, and a region mask is created for each preset appearance region in standard image coordinates. For the j-th preset appearance region, a sampling line is set every 4 pixels along the normal direction of the region boundary within the region mask. Each sampling line is 32 pixels long and located 16 pixels on each side of the boundary. The brightness value of each pixel on the sampling line is denoted as L(x), and the absolute value of the brightness difference between adjacent pixels is used as the boundary gradient value, i.e. The location where g(x) reaches its maximum value is taken as the standard boundary transition position corresponding to the sampling line. The dispersion of the standard boundary transition positions corresponding to all sampling lines within the same preset appearance area is calculated. The dispersion is the square root of the sum of the squared distances from each standard boundary transition position to the average position, divided by the number of sampling lines. After sorting the dispersion from low to high, the top 20% of the preset appearance areas are selected as stable boundary positioning areas; when the number of preset appearance areas is less than 10, at least 2 stable boundary positioning areas are selected.

[0041] For gloss-sensitive positioning areas, standard brightness responses under different light source directions are read within each preset appearance area. The gloss response difference of the j-th preset appearance area is the maximum value minus the minimum value of the average brightness of that area under all light source directions, divided by the sum of the average value of the average brightness under all light source directions and 1. After sorting the gloss response differences from high to low, the top 20% of the preset appearance areas are selected as gloss-sensitive positioning areas; when the number of preset appearance areas is less than 10, at least 2 gloss-sensitive positioning areas are selected. The resulting boundary stable positioning areas are used for planar attitude correction, and the gloss-sensitive positioning areas are used for curved surface tilt offset correction. Both are derived from standard appearance feature templates, avoiding reliance on occasional defects on the product surface for positioning.

[0042] The product to be tested is placed in the same positioning fixture as when the standard appearance image set was acquired. The image set is then acquired according to the light source number, light source direction, shooting angle, exposure time, and gain value recorded in the standard appearance image set. After acquisition, sampling lines are reset along the normal direction of the standard boundary within the stable positioning area. The positions of the sampling lines are given by the standard appearance feature template. The measured brightness value L′(x) for each sampling line is calculated. ,Will The location where the maximum value is reached is recorded as the transition position of the boundary to be measured. Then, within the gloss-sensitive positioning area, pixels whose brightness values ​​under the same light source direction exceed the average brightness value of the area plus 1.5 times the brightness standard deviation are identified as specular pixels. The brightness-weighted average position of all specular pixel coordinates is calculated to obtain the specular response position to be measured. The x-coordinate of the brightness-weighted average position is equal to the sum of the products of the x-coordinate and brightness value of each specular pixel divided by the sum of the brightness values ​​of each specular pixel. The y-coordinate is calculated using the same method.

[0043] A one-to-one correspondence is established between the transition positions of the boundary to be tested and the standard boundary transition positions in the standard appearance feature template. Let the i-th transition position of the boundary to be tested be pi, and the corresponding standard boundary transition position be qi. The planar attitude offset includes the rotation angle θ, the lateral translation a, and the longitudinal translation b. Let the rotated position to be tested be R(θ)pi+(a,b), where R(θ) is a two-dimensional rotation matrix. The first row of the matrix contains cosθ and -sinθ, and the second row contains sinθ and cosθ. θ, a, and b are obtained by minimizing the sum of squared distances between all corresponding points, even if... To minimize the error, the calculation first finds the center points of the transition position set of the test boundary and the standard boundary transition position set. Then, using the center points as a reference, the covariance matrix is ​​calculated. Singular value decomposition is performed on the covariance matrix to obtain the rotation angle θ. Subsequently, the difference between the two center points is used to obtain a and b. Based on θ, a, and b, each image in the test image set is mapped in reverse coordinates to obtain the test image set after preliminary attitude correction.

[0044] In the set of images to be tested after preliminary attitude correction, the position of the specular response to be tested within the gloss-sensitive localization region is recalculated and matched with the standard specular response position in the standard gloss features. For the r-th gloss-sensitive localization region, the displacement vector between the position of the specular response to be tested and the standard specular response position is calculated, and then this displacement vector is projected onto the arc normal direction of the region to obtain the specular residual displacement. Using the center coordinates of each gloss-sensitive localization region as the independent variable and the specular residual displacement as the dependent variable, a first-order tilted plane is fitted, which is expressed as... c0, c1, and c2 are obtained by minimizing the sum of squares of the differences between the residual specular displacement and d(x,y) within all gloss-sensitive localization regions. d(x,y) is used as the arc-surface tilt offset correction amount to perform coordinate compensation on the pre-corrected image set, ensuring that the arc-surface specular position corresponds to the standard gloss features. Finally, according to the region masks of each preset appearance region in the standard appearance feature template, the image set after arc-surface tilt offset correction is cropped, rotated, aligned, and resized to obtain a target region image set corresponding one-to-one with each preset appearance region.

[0045] The target region image set is input into the appearance feature extraction network to obtain the target appearance feature template. Based on the standard appearance feature template and the target appearance feature template, the consistency offset corresponding to each appearance region is calculated. When the consistency offset exceeds the threshold of the corresponding region, the abnormal appearance region and its offset feature are determined, forming an abnormal appearance feature group.

[0046] In some embodiments, target region images are sequentially read from the target region image set according to the arrangement order of the preset appearance regions in the standard appearance feature template. The j-th target region image is denoted as Ij, where j is an integer from 1 to m. Ij is scaled to the same pixel size as the j-th preset appearance region in the standard appearance feature template, and the same region mask is used to retain effective pixels. The appearance feature extraction network is formed by sharing convolution calculation and cascading four feature branches. The l-th layer shared convolution calculation is as follows: Where Al represents the output feature map of the l-th layer, Kl represents the convolution kernel, bl represents the bias, * represents the convolution operation, and ReLU represents the activation operation that sets negative values ​​to 0 and keeps non-negative values ​​unchanged. The region feature maps calculated by the shared convolution are respectively fed into the color branch, texture branch, gloss branch, and boundary transition branch, so that the target color features, target texture features, target gloss features, and target boundary transition features are consistent with the corresponding feature dimensions in the standard appearance feature template.

[0047] The color branch first converts Ij from the RGB color space to the Lab color space, calculates the mean, mean, and mean of the L channel, the variance of the a channel, and the variance of the b channel within the effective pixels, and statistically analyzes the color distribution in 16 equal-width intervals for the L channel, then concatenates these intervals to form the target color feature. The texture branch calculates the local gradient response in four directions on the shared convolutional feature map, and statistically analyzes the mean and variance of the response in each direction. Simultaneously, it extracts the global average pooling result from the last convolutional feature map and concatenates these intervals to form the target texture feature. The gloss branch reads the brightness response of the same preset appearance area under different light source directions, calculates the maximum brightness, the mean brightness, the proportion of highlight pixel area, and the highlight response position, forming the target gloss feature. The boundary transition branch reads the brightness curve along the normal direction of the region boundary, calculates the maximum gradient position, the transition band width, and the brightness difference between the two sides of the boundary, forming the target boundary transition feature. After all preset appearance areas have been calculated, the target appearance feature template is obtained.

[0048] The target appearance feature template is compared with the standard appearance feature template within the same region. In the j-th preset appearance region, the target color feature is denoted as Cxj, the standard color feature is denoted as C0j, and the color difference feature DCj is obtained by summing the absolute values ​​of the differences in each dimension. Ncj represents the number of dimensions for the color features. The target texture feature is denoted as Pxj, the standard texture feature as P0j, and the texture difference feature DPj is 1 minus the cosine of the angle between them. The cosine of the angle is equal to the inner product of Pxj and P0j divided by the product of the lengths of Pxj and P0j. The target gloss feature is denoted as Gxj, the standard gloss feature as G0j, and the gloss difference feature DGj is obtained by weighting the average absolute value of the brightness response difference under each light source direction with the distance to the specular response position. The weight for the brightness response difference is 0.7, and the weight for the specular response position distance is 0.3. The target boundary transition feature is denoted as Bxj, the standard boundary transition feature as B0j, and the boundary transition difference feature DBj is obtained by averaging the maximum gradient position difference, the transition band width difference, and the brightness difference on both sides of the boundary.

[0049] To eliminate the shared offset across the entire image set caused by slight fluctuations in camera exposure or light source attenuation, acquisition drift features are first extracted from the color and gloss differences of all preset appearance regions. For each color feature dimension, the signed difference between the target color feature and the standard color feature is calculated. The m signed differences are sorted in ascending order, and after removing the first and last 20% of the differences, the remaining 60% are averaged to obtain the color acquisition drift for that dimension. The gloss acquisition drift is obtained from the signed gloss differences of each preset appearance region in the same way. Then, the color acquisition drift is subtracted from the signed color differences of the corresponding dimension to obtain the corrected color difference feature; the gloss acquisition drift is subtracted from the signed gloss differences of the corresponding dimension to obtain the corrected gloss difference feature. This subtraction method preserves the abnormal offset in local areas while removing the acquisition error shared across the entire image set.

[0050] The corrected color difference features, corrected gloss difference features, texture difference features, and boundary transition difference features are fused to obtain the consistency offset Sj of the j-th preset appearance region. Where DCj′ represents the corrected color difference feature and DGj′ represents the corrected gloss difference feature. Preferably, wCj, wGj, wPj, and wBj are set to 0.3, 0.25, 0.25, and 0.2, respectively. When the number of standard good products reaches 50, each weight is obtained by normalizing the inverse of the variance of the corresponding feature in the standard good products, thus increasing the proportion of stable features in the fusion calculation.

[0051] For each preset appearance area, a region judgment limit Uj is set. Preferably, 50 standard good products are used to calculate the mean μj and standard deviation σj of the consistency offset of the j-th preset appearance area, and let... When σj is less than 0.01, it is used in the calculation as 0.01 to avoid the region judgment limit losing tolerance due to excessive sample concentration. Sj is compared with Uj; when Sj exceeds Uj, the j-th preset appearance region is determined as an abnormal appearance region. The offset percentage of four types of differences—color, gloss, texture, and boundary transition—to Sj is then calculated. The percentage equals the weighted difference of that type divided by Sj. The difference feature ranked first in percentage is determined as the offset feature; when the difference between the second and first ranked percentages does not exceed 0.05, the second ranked difference feature is also determined as an offset feature. Abnormal appearance regions, corresponding consistency offsets, and offset features together form an abnormal appearance feature group.

[0052] Based on the abnormal appearance feature group and production process data, the abnormal contribution degree of each candidate process node is generated, and the candidate process node with the largest abnormal contribution degree and its corresponding equipment number, fixture number, slot number and process parameters are used as the quality traceability result.

[0053] In some embodiments, abnormal appearance regions, offset features, and consistency offsets are read from the abnormal appearance feature group, and candidate process nodes are filtered in the production process data according to the offset feature category. When the offset feature is a color difference feature, the candidate process node includes the process number that affects color formation; when the offset feature is a texture difference feature, the candidate process node includes the process number that affects the surface roughness; when the offset feature is a gloss difference feature, the candidate process node includes the process number that affects the reflection; when the offset feature is a boundary transition difference feature, the candidate process node includes the process number that affects the boundary state of the region. Each candidate process node is denoted as Pn, where n is an integer from 1 to N. Pn corresponds to one process number and carries the equipment number, fixture number, slot number, and process parameter record segment.

[0054] Based on the entry and exit times of the product to be tested in Pn, a valid parameter segment is extracted from the process parameter recording segment. The r-th process parameter of the n-th candidate process node has H sampled values ​​within the valid parameter segment, denoted as zn,r,1 to zn,r,H. The target value for this process parameter is denoted as zn,r,0, the lower allowable limit is denoted as zn,r,L, and the upper allowable limit is denoted as zn,r,U. First, the segment mean mn,r of this process parameter is calculated, where mn,r is equal to the sum of the H sampled values ​​divided by H; then, the normalized deviation an,r is calculated. When zn,r,U is equal to zn,r,L, the denominator is set to 1 to avoid interrupting the calculation. The direction of segment change dn,r is determined by the difference between mn,r and zn,r,0. When mn,r is greater than zn,r,0, dn,r is set to 1; when mn,r is less than zn,r,0, dn,r is set to -1; and when the two are equal, dn,r is set to 0.

[0055] The region code of the abnormal appearance area in the product coordinate system is matched with the corresponding orientation codes of the equipment number, fixture number, and slot number in Pn. The region code is provided by the standard appearance feature template, and the orientation code is obtained from the fixture installation direction, slot number, and equipment operating surface recorded in the production process data. When the region code and the orientation code match, the orientation matching item bn,k is 1; when they do not match, bn,k is 0, where k represents the k-th abnormal appearance area. This method directly uses the relationship between the product coordinates and the slot for positioning, without relying on manual experience.

[0056] First, establish a correlation polarity table between offset features and the direction of change in process parameters. A correlation polarity of 1 indicates that an increase in process parameters will cause the corresponding offset feature to change in the positive direction; -1 indicates that an increase in process parameters will cause the corresponding offset feature to change in the negative direction; and 0 indicates that the two are not correlated. Taking dyeing time as an example, the correlation polarity of dyeing time on color deepening is 1; taking sandblasting pressure as an example, the correlation polarity of sandblasting pressure on texture coarsening is 1; and taking drying temperature as an example, the correlation polarity of drying temperature on gloss reduction is -1. The offset direction of the k-th abnormal appearance region is denoted as ek, with a positive offset of 1 and a negative offset of -1. If the product of the correlation polarity, dn,r, and ek is 1, then the direction matching term cn,r,k is 1; otherwise, cn,r,k is 0.

[0057] When merging the contribution results corresponding to equipment number, fixture number, and slot number under the same candidate process node, first calculate the single-area contribution value Qn,k of the nth candidate process node for the kth abnormal appearance area: Qn,k=Sk×Σ(an,r×cn,r,k×λn,r,k)×wn,k, where Sk is the consistency offset of the kth abnormal appearance area, λn,r,k is the absolute value of the correlation polarity, and wn,k is the orientation weight. When bn,k is 1 and cn,r,k is 1, Pn is the main contributing node and wn,k is 1; when bn,k is 0 and cn,r,k is 1, Pn is the auxiliary contributing node and wn,k is 0.45; when cn,r,k is not 1, wn,k is 0. The abnormal contribution degree Cn of the nth candidate process node is the sum of the single-area contribution values ​​of all abnormal appearance areas, i.e., Cn=ΣQn,k, and the summation range is all abnormal appearance areas within the abnormal appearance feature group.

[0058] The anomaly contribution of all candidate process nodes is sorted from highest to lowest value, and the candidate process node ranked first is selected as the target anomaly process node. The quality traceability results include the process number, equipment number, fixture number, slot number, names of process parameters involved in the calculation, mean of effective parameter segments, normalized deviation, and anomaly contribution of the target anomaly process node. In this method, the regional location, offset characteristics, and consistency offset in the anomaly appearance feature group all participate in the traceability calculation, enabling the localization of local appearance consistency anomalies to the corresponding process node and its equipment, fixture, and slot.

[0059] Example 2: To verify the appearance consistency detection and quality traceability effects achieved by the present invention, an aluminum alloy smart wearable product frame of the same model, color specification, and micro-arc oxidation gradient color process was selected as the verification object. This product includes a flat decorative area, a curved chamfered area, a high-gloss edge area, a hole perimeter area, a boundary transition area, and a shading edge area. The main appearance anomalies are slight local color differences, uneven curvature surface fogging, high-gloss dark spots, and whitening of the shading edges. These anomalies are not easily and stably displayed under direct light, but local consistency shifts occur under side light and low-angle grazing light, making it suitable for verifying the present invention's detection capability for complex appearance areas.

[0060] In this embodiment, 300 products were selected as verification samples. Among them, 240 samples were confirmed to be of acceptable appearance consistency after manual review, and 60 samples were found to be abnormal in appearance consistency. Of the 60 abnormal samples, 18 had color deviation, 14 had texture deviation, 16 had gloss deviation, and 12 had boundary transition deviation. For each product, frontal lighting images, left oblique lighting images, right oblique lighting images, top grazing lighting images, and bottom grazing lighting images were collected, resulting in a total of 1500 images for testing. The production process data includes the process numbers for six steps: sandblasting, anodizing, dyeing, sealing, cleaning, and drying, as well as the corresponding equipment numbers, fixture numbers, tank numbers, and process parameters for each step.

[0061] To establish a standard appearance feature template, 50 samples with acceptable appearance consistency were selected as standard good products. Their standard appearance image set was input into the appearance feature extraction network. Standard color features, texture features, gloss features, and boundary transition features were extracted according to six preset appearance regions. The mean and standard deviation of the consistency offset for each region were calculated to obtain the region judgment limit for each preset appearance region. The judgment limits for each preset appearance region are shown in Table 1.

[0062] Table 1. Statistical Results of Consistency Offset for Each Preset Appearance Area

[0063] Flat decorative area 0.118 0.021 0.181 Chamfered area of ​​curved surface 0.136 0.026 0.214 Highlight edge area 0.149 0.031 0.242 Peripheral area of ​​the hole 0.127 0.024 0.199 Boundary transition zone 0.142 0.028 0.226 Shielding edge area 0.153 0.030 0.243

[0064] As shown in Table 1, during the detection process, this invention first performs pose correction and region registration on the image set to be tested based on the standard boundary transition features and standard gloss features in the standard appearance feature template. Then, the target region image set is input into the appearance feature extraction network to obtain the target appearance feature template. To demonstrate the effectiveness of pose correction and curved surface tilt offset correction, Comparative Example 1 is set up. Comparative Example 1 only uses a positioning fixture to complete image acquisition and does not perform secondary correction based on the boundary stable positioning region and the gloss-sensitive positioning region. The detection results are shown in Table 2.

[0065] Table 2 Comparison of Attitude Correction and Region Registration Results

[0066] Comparative Example 1 2.86 4.73 17 This invention 0.64 1.08 3

[0067] As shown in Table 2, this invention obtains the planar attitude offset through a boundary stable positioning region and corrects the curved surface tilt offset through a gloss-sensitive positioning region. This reduces the average registration error of the planar region from 2.86 pixels to 0.64 pixels, and the average highlight offset of the curved surface from 4.73 pixels to 1.08 pixels. These results demonstrate that this invention can maintain a stable correspondence between the preset appearance area in the image set and the standard appearance feature template, even when there are slight placement deviations and curved surface reflection drift, thus avoiding the misidentification of acquisition position errors as actual appearance anomalies.

[0068] To further verify the consistency detection effect, comparative examples 2 and 3 were set up. Comparative example 2 used a fixed image difference method for detection, that is, directly comparing the grayscale difference between the image to be tested and the standard image; comparative example 3 used a single deep learning classification method for detection, only outputting qualified or abnormal results, without performing partition offset fusion for color, texture, gloss, and boundary transition. This invention compares the target appearance feature template with the standard appearance feature template within the same region, and after deducting the acquisition drift features common to the entire set of images to be tested, the consistency offset is obtained through fusion. The detection results are shown in Table 3.

[0069] Table 3 Comparison of Appearance Consistency Inspection Results

[0070] Comparative Example 2 60 48 12 21 89.0% Comparative Example 3 60 53 7 14 93.0% This invention 60 58 2 5 97.7%

[0071] As shown in Table 3, this invention correctly identified 293 out of 300 products, achieving a detection accuracy of 97.7%; it detected 58 out of 60 abnormal samples, with 2 missed detections. Comparative Example 2, due to its direct use of image differences, is easily affected by illumination drift and pose deviations, resulting in 21 falsely judged qualified products. Comparative Example 3, while possessing deep learning recognition capabilities, lacked a partitioned standard appearance feature template and did not perform acquisition drift subtraction for color and gloss differences, leading to insufficient stability in recognizing slight boundary whitening and localized highlights and dark spots. The improvements of this invention do not rely solely on classification results but rather achieve detection through the fusion of preset appearance regions, four types of appearance features, acquisition drift subtraction, and consistency offset, thus enhancing the detection capability of weak appearance anomalies.

[0072] To verify the adaptability of the present invention to different anomaly types, the detection results of 60 anomaly samples were statistically analyzed according to the offset feature category, and the results are shown in Table 4.

[0073] Table 4 Detection results for different offset feature categories

[0074] Color difference characteristics 18 18 0 100.0% Texture difference features 14 13 1 92.9% Gloss Difference Characteristics 16 15 1 93.8% Boundary transition difference characteristics 12 12 0 100.0%

[0075] As shown in Table 4, the detection rate of color difference features and boundary transition difference features of the present invention reached 100.0%, and the detection rates of texture difference features and gloss difference features reached 92.9% and 93.8%, respectively. These results indicate that the standard appearance feature template established by the present invention can cover different types of appearance anomalies, and is applicable not only to obvious color differences, but also to local consistency shift scenarios such as curved surface fogging, highlight dark spots, and whitening of occluded edges.

[0076] To verify the effectiveness of quality traceability, the actual sources of anomalies were recorded in 60 abnormal samples, and the quality traceability results output by this invention were compared with the actual sources of anomalies. The actual sources of anomalies included deviations in the dyeing bath concentration, deviations in the sandblasting pressure, deviations in the drying air velocity, and deviations in the sealing temperature. The traceability results are shown in Table 5.

[0077] Table 5 Comparison of Quality Traceability Results

[0078] Concentration deviation in dyeing bath 18 18 17 16 Sandblasting process sandblasting pressure deviation 14 13 13 12 Drying process wind speed deviation 16 15 14 14 Temperature deviation during sealing process 12 12 11 11 total 60 58 55 53

[0079] As shown in Table 5, the present invention correctly traced 58 abnormal process numbers, with a process traceability accuracy rate of 96.7%; correctly traced 55 equipment numbers, with an equipment traceability accuracy rate of 91.7%; and correctly traced 53 fixture numbers and slot numbers, with a slot traceability accuracy rate of 88.3%. These results demonstrate that the present invention does not merely provide anomaly categories after detection, but rather combines the abnormal appearance area, offset characteristics, and consistency offset with the process numbers, equipment numbers, fixture numbers, slot numbers, and process parameters in the production process data to form the anomaly contribution of candidate process nodes. This allows for the localization of appearance consistency anomalies to specific production nodes.

[0080] To illustrate the improvement in screening efficiency brought about by this invention, the screening time was statistically analyzed for four batches of production records corresponding to the aforementioned 60 abnormal samples. The manual, experience-based screening method involved process personnel sequentially checking the sandblasting, anodizing, dyeing, sealing, cleaning, and drying processes, and retrieving equipment operation records for each. The method of this invention directly outputs the candidate process node with the highest contribution to the anomaly based on the quality traceability results, along with its corresponding equipment number, fixture number, slot number, and process parameters. The statistical results are shown in Table 6.

[0081] Table 6 Comparison of Anomaly Detection Efficiency

[0082] Manual experience screening 6 74 no This invention 1.4 11 yes

[0083] As shown in Table 6, when using manual experience for inspection, an average of 6 processes need to be checked, with an average inspection time of 74 minutes. After adopting the present invention, the average number of processes involved is reduced to 1.4, the average inspection time is reduced to 11 minutes, and the slot information can be directly output. Therefore, the present invention, through the continuous correlation between appearance inspection results and production process data, can reduce line stoppages for process-by-process re-inspection and improve the efficiency of anomaly location.

[0084] Based on the data in Tables 1 to 6, this embodiment demonstrates that the present invention, through standard appearance feature templates, posture correction, region registration, acquisition drift deduction, consistency offset fusion, and anomaly contribution tracing, achieves a product appearance consistency detection accuracy of 97.7% and an anomaly process tracing accuracy of 96.7%, while reducing the average investigation time from 74 minutes to 11 minutes. These effects directly stem from the present invention's technical means of regionalizing, directionalizing, and nodding-based association between local appearance features and production process data, which solves the problems of unstable identification of weak consistency anomalies and difficulty in locating the source of anomalies in existing appearance inspection methods.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A product appearance consistency detection and quality traceability method based on deep learning, characterized in that, include: Obtain a standard appearance image set and production process data corresponding to the product to be tested. The production process data includes process number, equipment number, fixture number, slot number and process parameters. The standard appearance image set is input into the appearance feature extraction network, and standard color features, texture features, gloss features and boundary transition features are extracted according to the preset appearance area to establish a standard appearance feature template. Collect a set of images of the product to be tested under the same light source direction and shooting angle, and perform posture correction and region registration on the set of images to be tested according to the standard appearance feature template to obtain the target area image set; The target region image set is input into the appearance feature extraction network to obtain the target appearance feature template. Based on the standard appearance feature template and the target appearance feature template, the consistency offset corresponding to each appearance region is calculated. When the consistency offset exceeds the threshold of the corresponding region, the abnormal appearance region and its offset feature are determined to form an abnormal appearance feature group. Based on the abnormal appearance feature group and production process data, the abnormal contribution degree of each candidate process node is generated, and the candidate process node with the largest abnormal contribution degree and its corresponding equipment number, fixture number, slot number and process parameters are used as the quality traceability result.

2. The product appearance consistency detection and quality traceability method based on deep learning according to claim 1, characterized in that, Obtain the standard appearance image set and production process data corresponding to the product to be inspected, including: Before the product to be inspected enters the appearance consistency inspection station, a standard appearance image set corresponding to a standard product of the same product model, color specification and appearance processing technology as the product to be inspected is obtained; the unique product identifier of the product to be inspected is read, and the corresponding process number, equipment number, fixture number, slot number and process parameters are obtained based on the unique product identifier; the standard appearance image set and production process data are bound according to the unique product identifier to form the basic inspection data corresponding to the product to be inspected.

3. The product appearance consistency detection and quality traceability method based on deep learning according to claim 2, characterized in that, Acquisition of the standard appearance image set includes: Place the standard product in the positioning fixture and align the reference edge, assembly hole or outer contour of the standard product with the detection coordinate system; acquire images of the standard product according to the preset shooting angle, and light sources in different directions are lit in sequence to obtain a set of standard appearance images covering the planar area, curved area, highlight edge area and boundary transition area; record the product model, color specification, light source number, light source angle, camera number, shooting angle, exposure time and gain value for each standard appearance image.

4. The product appearance consistency detection and quality traceability method based on deep learning according to claim 1, characterized in that, Establish standard appearance feature templates, including: The standard appearance image set is processed by image denoising, white balance correction, distortion correction, brightness normalization, and image size unification. Based on the product structural characteristics, the appearance surface is divided into planar areas, curved areas, highlight areas, hole perimeter areas, boundary transition areas, and occlusion edge areas, and a region mask is created for each preset appearance area. The standard color features, texture features, gloss features, and boundary transition features are regionalized and summarized according to the region mask to obtain the standard appearance feature template.

5. The product appearance consistency detection and quality traceability method based on deep learning according to claim 4, characterized in that, Extraction of standard color features, texture features, gloss features, and boundary transition features, including: The color feature extraction branch of the appearance feature extraction network extracts the color distribution features and color change curves within the preset appearance area; the texture feature extraction branch of the appearance feature extraction network extracts the sandblasting particle distribution features, local texture direction features, and deep texture embedding features within the preset appearance area; the gloss feature extraction branch and boundary transition feature extraction branch of the appearance feature extraction network extract the gloss response features within the preset appearance area and the boundary change features between adjacent appearance areas, respectively.

6. The product appearance consistency detection and quality traceability method based on deep learning according to claim 1, characterized in that, Pose correction and region registration are performed on the image set to be tested based on standard appearance feature templates, including: Call the standard boundary transition feature in the standard appearance feature template, and select the preset appearance region whose boundary transition position dispersion meets the preset sorting conditions as the stable boundary positioning region. Call the standard gloss feature in the standard appearance feature template, and select the preset appearance area whose brightness response difference under different light source directions meets the preset sorting conditions as the gloss-sensitive positioning area. The planar pose correction of the image set to be tested is performed using the boundary stable positioning region, and the arc surface tilt offset correction of the product to be tested is performed using the gloss-sensitive positioning region.

7. The product appearance consistency detection and quality traceability method based on deep learning according to claim 6, characterized in that, Planar pose correction of the image set under test is performed using a boundary-stabilized localization region, including: Within the stable positioning area of ​​the boundary, sampling lines are set along the direction of the standard boundary normal, and the transition position of the boundary to be tested corresponding to each sampling line is extracted from the set of images to be tested. Establish a correspondence between the transition position of the boundary to be tested and the standard boundary transition position in the standard appearance feature template to obtain the rotation offset and translation offset of the image set to be tested relative to the standard appearance image set. The image set under test is mapped in reverse coordinates based on rotation offset and translation offset to obtain the image set under test after preliminary attitude correction.

8. The product appearance consistency detection and quality traceability method based on deep learning according to claim 7, characterized in that, Correcting the curved surface tilt offset of the product under test using a gloss-sensitive positioning area includes: The location of the specular response to be tested within the gloss-sensitive localization area is extracted from the image set after preliminary attitude correction. The specular response position to be tested is matched with the standard specular response position in the standard appearance feature template to obtain the specular residual displacement corresponding to the gloss-sensitive positioning area. Based on the specular residual displacement, coordinate compensation is performed on the preliminary attitude-corrected image set, and the image set of the target region is obtained by cropping, rotating and aligning, and unifying the size according to the region mask in the standard appearance feature template.

9. The product appearance consistency detection and quality traceability method based on deep learning according to claim 1, characterized in that, A group of abnormal appearance features is formed, including: The target region image set is input into the appearance feature extraction network according to the preset appearance region order in the standard appearance feature template to extract target color features, target texture features, target gloss features and target boundary transition features that are consistent with the dimensions of the standard appearance feature template. The target appearance feature template is compared with the standard appearance feature template in the same area to obtain the color difference features, texture difference features, gloss difference features and boundary transition difference features corresponding to each preset appearance area. After subtracting the acquisition drift features common to the entire set of images under test from the color difference features and gloss difference features, the consistency offset is obtained by fusing them, and the abnormal appearance regions and their offset features are determined based on the consistency offset.

10. The product appearance consistency detection and quality traceability method based on deep learning according to claim 9, characterized in that, Anomaly contribution values ​​for each candidate process node are generated based on abnormal appearance feature groups and production process data, including: Candidate process nodes are selected from the production process data according to the offset feature category, and valid parameter fragments are extracted when the product to be tested actually passes through each candidate process node. The region code of the abnormal appearance area in the product coordinates is matched with the function orientation code of the equipment number, fixture number and slot number corresponding to the candidate process node to obtain the region response matching result; Based on the regional response matching results, the deviation status of the effective parameter segments, and the direction of change of the offset characteristics, the contribution results corresponding to the equipment number, fixture number, and slot number under the same candidate process node are merged to obtain the abnormal contribution degree of each candidate process node.

Images