Printing fat oil defect detection method and electronic equipment shell based on UNet model

Abstract

The application relates to the technical fields of artificial intelligence and printing, and provides a printing fat oil defect detection method based on a UNet model and an electronic equipment shell, wherein a practical UNet model capable of identifying different fat oil defects is trained, when the electronic equipment shell is produced, the ink printing layer of the current electronic equipment shell substrate is imaged to obtain an ink printing layer image, the ink printing layer image is input into the practical UNet model, the fat oil defect of the printing pattern of the ink printing layer of the current electronic equipment shell substrate is identified through the practical UNet model, and when the practical UNet model identifies that the ink printing layer of the current electronic equipment shell substrate has the fat oil defect of the printing pattern, the current electronic equipment shell substrate with the fat oil defect is transported to a defective product recycling area, so that when the fat oil defect of the printing pattern of the ink printing layer of the substrate is detected, the detection precision and the detection efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of artificial intelligence, computer vision, image recognition and printing, and particularly relates to a printing fat oil defect detection method based on a UNet model and an electronic device shell. BACKGROUND

[0002] In the printing production process of the electronic device shell, generally includes printing screen preparation, ink printing, standing, baking and film covering and other processes. After the printing screen is prepared, the configured ink is printed to the base color layer of the substrate (for example, the green film surface of the substrate) through the printing screen, and the ink printing layer is covered by the frosting film after standing and baking. In the process of printing the configured ink to the base color layer of the substrate through the printing screen, it is necessary to detect whether the pattern of the ink printing layer of the substrate has fat oil defects. In the prior art, in order to detect the fat oil defects of the pattern of the ink printing layer of the substrate, the detection is mainly carried out by visual detection technology. The detection result of visual detection is easily affected by the experience, vision and environmental light change of the operator, and the detection precision and efficiency are low.

[0003] In summary, in the printing production process of the existing electronic device shell, the detection technology of the fat oil defects of the pattern of the ink printing layer of the substrate has the technical problems of low detection precision and low detection efficiency. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a printing fat oil defect detection method based on a UNet model and an electronic device shell to improve the detection precision and efficiency of the fat oil defects of the pattern of the ink printing layer of the substrate.

[0005] In a first aspect, the present application provides a printing fat oil defect detection method based on a UNet model, comprising:

[0006] The UNet model is trained by using the image of the printing pattern containing fat oil defects to obtain a practical UNet model capable of identifying different fat oil defects, the fat oil defects include boundary blur defects and thickness uneven defects, the boundary blur defects represent that the ink of the printing pattern spreads to the area outside the predetermined pattern, resulting in blurred pattern edges, and the thickness uneven defects represent that the ink of the printing pattern in some areas is too much, forming a convex or abnormal thickness;

[0007] When the electronic device shell is produced, the ink printing layer of the current electronic device shell substrate after standing and baking is imaged by using a diffuse reflection light source camera system to obtain an image of the ink printing layer of the current electronic device shell substrate;

[0008] Inputting an image of an ink printing layer of a current electronic device housing substrate into the practical UNet model, and identifying fat defects in a printing pattern of the ink printing layer of the current electronic device housing substrate using the practical UNet model;

[0009] When the practical UNet model identifies that the ink printing layer of the current electronic device housing substrate has a fat oil defect in the printed pattern, the current electronic device housing substrate with the fat oil defect is transported to a defective product recycling area.

[0010] In a second aspect, the present invention provides an electronic device housing, which is inspected during a printing process using the above-mentioned UNet model-based method for detecting printing fat oil defects.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] The present invention provides a method for detecting printed fat defects based on a UNet model and an electronic device housing. The UNet model is trained by using an image of a printed pattern containing fat defects to obtain a practical UNet model capable of identifying different fat defects. When printing and producing the electronic device housing, the ink printing layer of the current electronic device housing substrate after standing and baking is imaged by a diffuse reflection light source camera system to obtain an ink printing layer image of the current electronic device housing substrate. The ink printing layer image of the current electronic device housing substrate is input into the practical UNet model. The fat defects in the printed pattern of the ink printing layer of the current electronic device housing substrate are identified by the practical UNet model. When the practical UNet model identifies that the ink printing layer of the current electronic device housing substrate has a fat defect in the printed pattern, the current electronic device housing substrate with the fat defect is transported to a defective product recovery area, thereby improving the detection accuracy and efficiency when detecting the fat defects in the ink printing layer pattern of the substrate. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings described herein are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0014] Figure 1 The figure is a flow chart of a method for detecting printing fat oil defects based on a UNet model according to an embodiment of the present invention.

[0015] Figure 2 This is a flow chart of training a UNet model using images of printed patterns containing fat defects in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0017] See also Figure 1 , Figure 2 The present invention provides a method for detecting printing fat oil defects based on a UNet model and an electronic device housing. The electronic device housing is detected using the method for detecting printing fat oil defects based on a UNet model during a printing production process. The method for detecting printing fat oil defects based on a UNet model includes:

[0018] Step S101: Training a UNet model using images of printed patterns containing oil defects to obtain a practical UNet model capable of identifying different oil defects. The oil defects include blurred boundary defects and uneven thickness defects. The blurred boundary defect indicates that ink in a printed pattern has diffused into areas outside the predetermined pattern, resulting in blurred pattern edges. The uneven thickness defect indicates that certain areas of the printed pattern have excessive ink, forming protrusions or abnormal thickness.

[0019] Step S102: When printing and producing the electronic device housing, the ink printing layer of the current electronic device housing substrate after standing and baking is imaged by a diffuse reflection light source camera system to obtain an image of the ink printing layer of the current electronic device housing substrate;

[0020] Step S103: Inputting the ink printing layer image of the current electronic device housing substrate into the practical UNet model, and identifying fat defects in the printing pattern of the ink printing layer of the current electronic device housing substrate using the practical UNet model;

[0021] Step S104: When the practical UNet model identifies that the ink printing layer of the current electronic device housing substrate has a fat oil defect in the printed pattern, the current electronic device housing substrate with the fat oil defect is transported to a defective product recycling area.

[0022] It should be noted that in the prior art, the detection of fat oil defects of printed patterns in electronic device shell ink printing mainly relies on manual visual detection. The human eye has limited ability to recognize small or complex defects, and is prone to misjudgment or omission. In addition, in a mass production environment, manual detection is time-consuming and prone to fatigue, and cannot meet the needs of efficient production. In addition, the detection results of manual detection are affected by the experience, subjective judgment and environmental light changes of the operator, and lack consistency. In the present embodiment, a UNet model is trained using images of printed patterns containing fat oil defects to obtain a practical UNet model capable of identifying different fat oil defects. During the printing of electronic device shells, the ink printing layer of the current electronic device shell substrate after standing and baking is imaged by a diffuse light source camera system to obtain an image of the ink printing layer of the current electronic device shell substrate. The image of the ink printing layer of the current electronic device shell substrate is input into the practical UNet model, and the fat oil defects of the printed pattern of the ink printing layer of the current electronic device shell substrate are identified by the practical UNet model. When the practical UNet model identifies that the ink printing layer of the current electronic device shell substrate has fat oil defects of the printed pattern, the current electronic device shell substrate with fat oil defects is transported to a defective product recycling area, thereby improving the detection accuracy and efficiency when detecting the fat oil defects of the pattern of the ink printing layer of the substrate.

[0023] It should be noted that the UNet model is widely used in the field of image processing due to its precise segmentation ability for image details, and is particularly suitable for processing complex fat oil defects (such as boundary blur defects and thickness unevenness defects) of printed patterns in the present embodiment. In the present embodiment, the UNet model is trained using sample images containing typical fat oil defects, so that the UNet model can efficiently identify and classify specific types of defects. When the practical UNet model identifies that the ink printing layer of the current electronic device shell substrate has fat oil defects of the printed pattern, the current electronic device shell substrate with fat oil defects is transported to a defective product recycling area, thereby dataizing and automating the detection process and reducing human intervention and errors.

[0024] It should be noted that in the diffuse light source camera system, the diffuse light source can provide stable lighting conditions, eliminate the influence of light angle and intensity changes on image quality, and ensure imaging consistency. The diffuse light source camera system can capture the ink pattern and its defects in detail, providing high-quality image data input for defect recognition by the practical UNet model.

[0025] In some preferred embodiments, when training the UNet model with images containing printed patterns with fat oil defects, the method comprises: collecting image samples of printed patterns containing boundary fuzzy defects and uneven thickness defects, and labeling the image samples to mark the areas with defects and the types of defects to obtain real labels of the training data; inputting the image samples into the UNet model to train the UNet model, and calculating a loss function by comparing the difference between the output predicted by the UNet model and the real labels, and optimizing the UNet model to obtain the practical UNet model. It should be noted that in order to enable the UNet model to effectively learn the characteristics of fat oil defects, image samples containing boundary fuzzy defects and uneven thickness defects are needed. These samples represent various defect conditions that may be encountered in actual production, and the UNet model can master how to distinguish different types of defects by learning these samples. Each image sample needs to correspond to a real label, and the label marks the defect area and its type. The UNet is trained based on the input image samples and their corresponding labels, and can clearly know which areas are defects and what type of defects they are. In this embodiment, by inputting the image samples into the UNet model, the UNet model can extract features (such as edges, textures, etc.) in the image through convolution layers, and gradually recover these features to finally predict the defect area in the image. The training of the UNet model is based on the error backpropagation mechanism, and the loss function is used to measure the difference between the output predicted by the model and the real label. By calculating this difference, the UNet model can know which areas are not accurately identified, so as to gradually improve the recognition accuracy by optimizing and adjusting its internal parameters. Preferably, the loss function adopts cross-entropy loss or Dice coefficient loss. Cross-entropy loss or Dice coefficient loss can help the UNet model to effectively optimize and improve its segmentation ability of defect areas. In addition, in this embodiment, through multiple training and optimization, the UNet model will gradually improve its recognition ability of fat oil defects, so that the model has high accuracy and robustness in actual application. The specific optimization process can be to gradually adjust the model parameters (such as convolution kernel weight, network depth, etc.) to make it better adapt to the needs of the task, for example, in actual production, to identify and accurately classify boundary fuzzy defects and uneven thickness defects.

[0026] In some preferred embodiments, when labeling the image sample, the image sample is denoised using a Gaussian filter algorithm, a median filter algorithm or a mean filter algorithm to remove random noise of the image in the image sample, to obtain a random noise removed image sample; and a histogram equalization algorithm or an adaptive histogram equalization algorithm is used to enhance the contrast of the random noise removed image sample, to highlight the boundary blur defect and the thickness uneven defect, to obtain a defect highlighted image sample. It should be noted that Gaussian filtering is a linear filter that suppresses high-frequency noise (such as random noise) in the image by weighted averaging. It is very effective for removing Gaussian noise and smoothing details in the image. After applying Gaussian filtering, the noise of the image is smoothed, which can eliminate random interference and highlight the real image structure and defect area. Median filtering is a nonlinear filtering method, which is particularly suitable for removing salt and pepper noise (bright or dark spots appearing in the image). Compared with Gaussian filtering, median filtering can better preserve the edge features of the image and avoid blurred boundaries. Mean filtering is a simple linear filtering method that performs local averaging on the image to remove noise in the image. It has good effect on removing uniformly distributed noise, but may lose some details of the image. Noise in the image may mask or distort the actual defect area, making the labeling inaccurate. Denoising can effectively remove these random noises and preserve important image details (such as the outline and texture of the defect), thereby improving the accuracy of subsequent defect area recognition. Histogram equalization adjusts the gray scale distribution of the image to make the contrast of the image more uniform, especially between the dark and bright parts of the image, enhancing the details of low-contrast images and making the boundary blur and thickness uneven defects more easily identifiable. Unlike traditional histogram equalization, adaptive histogram equalization (CLAHE) enhances contrast in local regions, which is particularly effective for images with large-scale brightness variations. It can effectively avoid over-enhancement caused by traditional methods and maintain the clarity of image details. In some images, especially complex printed patterns, the contrast between the defect area and the background may be low, making it difficult to detect defects. By increasing the contrast of the image, especially highlighting the boundary blur and thickness uneven defects, the visibility of the defect area can be significantly improved, providing a clear basis for subsequent defect area and type labeling. It should be noted that boundary blur defects are usually characterized by unclear or blurred boundaries of the pattern. Through denoising and contrast enhancement, the blurred edges can become more obvious, facilitating accurate segmentation by subsequent edge detection algorithms. Thickness uneven defects often cause uneven ink accumulation, forming raised or thickness abnormal areas. Through image enhancement, the thickness uneven areas can be more easily separated from the background, thereby facilitating the labeling of these abnormal areas.Overall, the de-noising can eliminate interference factors, and the contrast enhancement can highlight the characteristics of defects, making the boundary fuzzy defects and thickness uneven defects more obvious, thereby facilitating the subsequent labeling process. This process can ensure the accuracy of the labeling results and improve the efficiency and accuracy of the labeling.

[0027] In some preferred embodiments, after obtaining the defect highlighted image sample, an edge detection algorithm is used to identify the defect edges of the image in the defect highlighted image sample to obtain a potential defect region with a preliminary contour; a region growing algorithm is used to expand the potential defect region, automatically marking out a defect region with a fuzzy boundary and a defect region with uneven thickness, and automatically marking out a fuzzy boundary defect type and an uneven thickness defect type according to the region morphology, edge features and thickness variation of the potential defect region. It should be noted that the main role of edge detection is to identify regions with large changes in brightness in the image, which usually represent the boundaries or structures in the image. For defects in the printed pattern, especially fuzzy boundary defects and uneven thickness defects, the edges of the defects in the image usually exhibit features different from the background. In this embodiment, the edge detection algorithm (such as Canny edge detection or Sobel operator) can capture the gray scale variation in the image, identify these fuzzy edges, and thus highlight the potential defect region. In addition, uneven thickness defects usually exhibit excessive ink accumulation in some regions, resulting in large changes in the thickness of the pattern. The edge detection algorithm can identify the edges of these regions, and when the ink accumulates or is unevenly distributed, the gray scale variation in the image also exhibits obvious edge features. In addition, the region growing algorithm is a region expansion algorithm based on seed points, which can start from a seed point and gradually expand the region according to certain similarity criteria (such as gray scale value, texture feature) until the stopping criteria are met. This algorithm can automatically segment the region and mark out the corresponding defect region. Fuzzy boundary defects usually have fuzzy edges, which are not easy to accurately define their range directly using traditional edge detection methods. The region growing algorithm can further expand these edges based on edge detection, thus accurately marking out the complete range of the fuzzy defect region. For uneven thickness defects, the region growing algorithm can identify the regions with uneven ink distribution according to the gray scale distribution or texture features in the image, and expand these regions to finally mark out the defect region. Regions with large thickness variations usually exhibit large fluctuations in the gray scale image, and the region growing algorithm can accurately identify these regions. In addition, after marking out the potential defect region, judging the defect type is a crucial step. Fuzzy boundary defects and uneven thickness defects exhibit different features in the image, so they can be distinguished by region morphology, edge features and thickness variation. Fuzzy boundary defects usually exhibit unclear or fuzzy pattern edges. After obtaining the preliminary contour by the edge detection algorithm, the edges of the defect region may exhibit large fuzziness, usually with irregular morphology, and the region morphology is fuzzy and has no obvious clear boundary. Uneven thickness defects exhibit abnormal changes in ink accumulation or thickness. By analyzing the gray scale value or texture information in the image, the regions with excessive ink can be identified, and usually these regions have large gray scale variation and exhibit a relatively convex feature on the surface.The process of automatically marking the defect type can be based on analysis of the region morphology (such as the shape, size, distribution of the region), edge features (such as edge sharpness, smoothness), and thickness variation (such as gray scale distribution, texture features), thereby efficiently identifying and marking the defect region, automatically determining the type of defect, and greatly improving the marking efficiency and accuracy.

[0028] In some preferred embodiments, the automatically labeled boundary-ambiguous defect regions and the thickness-uneven defect regions are compared with the actual defect regions to correct the automatically labeled boundary-ambiguous defect regions and the thickness-uneven defect regions, and the automatically labeled boundary-ambiguous defect types and the thickness-uneven defect types are compared with the actual defect types to correct the automatically labeled boundary-ambiguous defect types and the thickness-uneven defect types, to obtain the true labels of the training data. It should be noted that although edge detection and region growing algorithms can effectively label defect regions, due to image quality, background complexity or limitations of the algorithm, there may be certain errors in the automatically labeled defect regions, such as mislabeling, missing labeling or incomplete labeling. Comparing the automatically labeled defect regions with the actual defect regions can help identify and correct these errors, ensuring the accuracy of the training data. Through such comparison, inaccurate labeling due to algorithm limitations can be eliminated, ensuring the completeness and accuracy of the defect regions. In addition, for boundary-ambiguous and thickness-uneven defects, the automatic labeling algorithm identifies the defect type through image features such as edge features, region morphology and gray level changes. However, the automatically labeled defect type may be affected by background noise, sensitivity settings of the algorithm or complexity of the defect appearance, resulting in misjudgment. In this embodiment, by comparing the automatically labeled defect type with the actual defect type, these misjudgments can be identified and corrected. For example, some ambiguous boundaries may be incorrectly labeled as thickness uneven, or vice versa, some ink layer abnormalities may be mistaken for boundary ambiguity. Through comparison, it can be ensured that the type of defect is correctly identified and adjusted to more accurately reflect the actual situation. Specifically, if there is a difference between the automatically labeled defect type and the actual defect type, the morphological features of the region (such as shape, symmetry, size, etc.) can be analyzed to further adjust and correct. For example, boundary-ambiguous defects often exhibit irregular shapes, while thickness-uneven defects usually exhibit local thickness variations. In addition, thickness-uneven defects often exhibit certain texture changes in the image, and by comparing with the actual defect, the features of these texture changes can be identified to determine whether they belong to thickness-uneven defects. For boundary-ambiguous defects, the sharpness of the image edge is compared to determine whether it belongs to boundary-ambiguous defects. In this embodiment, by comparing and correcting the automatically labeled defect regions and types, the quality of the training data can be effectively improved, ensuring that it reflects the actual defect situation. The true label is crucial for training the model, because the accuracy and generalization ability of the model are directly affected by the quality of the training data. By correcting the automatically labeled defect regions and defect types, more diverse and high-quality training data can be obtained, which helps the model to perform more stably and accurately in the recognition of different types of defects.

[0029] In further embodiments, after the current electronic device shell substrate with the fat oil defect is transported to the defective product recycling area, if the fat oil defect identified by the practical UNet model is a boundary fuzzy defect, the edges of the fuzzy area are compared with the edges of the normal area to determine the degree of fuzziness; if the fuzzy edge is obvious, it is determined that the boundary fuzzy defect is caused by uneven printing pressure or ink diffusion, prompting to adjust the printing pressure and printing speed to ensure uniform ink coating, or to improve the ink formula to reduce ink diffusion; if the boundary of the fuzzy area is relatively regular, it is determined that the boundary fuzzy defect is caused by too low ink viscosity or too fast solvent evaporation, prompting to increase the proportion of viscous components of the ink to improve the viscosity of the ink, or to increase the solidification speed of the ink to prevent too fast solvent evaporation; if the fat oil defect identified by the practical UNet model is a thickness uneven defect, the ink layer distribution of the thickness uneven area is compared to determine the thickness difference; if the ink layer of some areas is obviously too thick, it is determined that the thickness uneven defect is caused by uneven ink coating or too large printing pressure, prompting to adjust the uniformity of the ink coating equipment or optimize the printing pressure to ensure uniform ink coating; if the ink layer of some areas is obviously too thin, it is determined that the thickness uneven defect is caused by insufficient ink supply or too small printing pressure, prompting to increase the amount of ink supply or adjust the printing pressure to ensure uniform ink layer coating and avoid local ink accumulation or deficiency. It should be noted that the causes of boundary fuzzy defects are usually related to ink diffusion or uneven printing pressure. In this embodiment, by comparing the edges of the fuzzy area with the edges of the normal area to determine the degree of fuzziness, different defect types can be effectively distinguished to help determine the cause of the defect, so as to accurately adjust the related process parameters or ink formula, reduce subsequent manual inspection and rework, and improve production efficiency. In addition, the thickness uneven defect mainly manifests as local area ink coating being too thick or too thin. In this embodiment, by comparing the thickness difference and judging the characteristics of the problem area, the specific cause of the defect can be determined to provide an accurate solution.

[0030] In further embodiments, the utility UNet model first identifies the boundary fuzzy defect when identifying the fat oil defect, and then identifies the uneven thickness defect when the boundary fuzzy defect does not exist. It should be noted that the boundary fuzzy defect is usually manifested as the blurring or expansion of the edge of the printed pattern, and the area involved is usually the edge of the pattern, the morphology and characteristics are more obvious, and in most cases, the blurred edge is usually significantly contrasted with the normal area, and is easily identified by edge detection or image segmentation algorithm. The uneven thickness defect is manifested as the thickness difference of the ink layer, and such difference is usually local and more hidden in the image, and needs more accurate detection and identification to determine. In the fat oil defect detection of the present embodiment, the easier-to-detect defect type is identified first, and then the more difficult-to-detect defect type is identified, for example, the boundary fuzzy defect is identified first, and then the uneven thickness defect is identified when the boundary fuzzy defect does not exist, so that the identification efficiency can be improved and the computing resources can be saved.

[0031] It should be noted that the above embodiments are only preferred specific embodiments of the present application, and the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for detecting defects of printed fat based on a UNet model, characterized in that, The application comprises the following steps: training a UNet model by using images containing printing patterns with excess ink defects, so as to obtain a practical UNet model capable of identifying different excess ink defects, wherein the excess ink defects include a boundary blur defect and a thickness uneven defect, the boundary blur defect represents that ink of a printing pattern spreads to an area outside a predetermined pattern, resulting in a blurred pattern edge, and the thickness uneven defect represents that some areas of the printing pattern have excessive ink, forming a protrusion or abnormal thickness; during the production of an electronic device shell, imaging the ink printing layer of the current electronic device shell substrate after standing and baking by using a diffuse reflection light source camera system, so as to obtain an image of the ink printing layer of the current electronic device shell substrate; inputting the image of the ink printing layer of the current electronic device shell substrate into the practical UNet model, and identifying the excess ink defects of the printing pattern of the ink printing layer of the current electronic device shell substrate by using the practical UNet model; when the practical UNet model identifies that the ink printing layer of the current electronic device shell substrate has excess ink defects of the printing pattern, transporting the current electronic device shell substrate with excess ink defects to a defective product recycling area; after the current electronic device shell substrate with excess ink defects is transported to the defective product recycling area, if the excess ink defects identified by the practical UNet model are the boundary blur defect, comparing the edge of the blurred area with the edge of a normal area to determine the blur degree; if the blurred edge is obvious, it is determined that the boundary blur defect is caused by uneven printing pressure or ink spreading, and it is prompted to adjust the printing pressure and printing speed to ensure uniform ink coating, or to improve the ink formula to reduce ink spreading; if the boundary of the blurred area is relatively regular, it is determined that the boundary blur defect is caused by excessively low ink viscosity or excessively fast solvent evaporation, and it is prompted to increase the proportion of viscous components of the ink to improve the viscosity of the ink, or to increase the solidification speed of the ink to prevent excessive solvent evaporation; if the excess ink defects identified by the practical UNet model are the thickness uneven defect, comparing the ink layer distribution of the thickness uneven area to determine the thickness difference; if the ink layer of some areas is obviously too thick, it is determined that the thickness uneven defect is caused by uneven ink coating or excessively large printing pressure, and it is prompted to adjust the uniformity of the ink coating equipment or optimize the printing pressure to ensure uniform ink coating; if the ink layer of some areas is obviously too thin, it is determined that the thickness uneven defect is caused by insufficient ink supply or excessively small printing pressure, and it is prompted to increase the amount of ink supply or adjust the printing pressure to ensure uniform ink layer coating and avoid local ink accumulation or deficiency; when identifying the excess ink defects, the practical UNet model first identifies the boundary blur defect, and then identifies the thickness uneven defect when the boundary blur defect does not exist. 2.The UNet model-based printed fat oil defect detection method of claim 1, wherein, The UNet model is trained by using images of printed patterns containing fat oil defects, including: collecting image samples of printed patterns containing boundary fuzzy defects and uneven thickness defects, and labeling the image samples to mark the areas where defects exist and the types of defects, to obtain real labels of training data; inputting the image samples into the UNet model to train the UNet model, and calculating a loss function and optimizing the UNet model by comparing the difference between the output predicted by the UNet model and the real labels, to obtain the practical UNet model. 3.The UNet model-based printed fat oil defect detection method of claim 2, wherein, The loss function adopts cross-entropy loss or Dice coefficient loss. 4.The UNet model-based printed fat oil defect detection method of claim 2, wherein, When labeling the image samples, the image samples are denoised to remove random noise of the images in the image samples, to obtain image samples after random noise removal; and the image samples after random noise removal are subjected to contrast enhancement to highlight the boundary fuzzy defects and the uneven thickness defects, to obtain defect-highlighted image samples. 5.The UNet model-based printed fat oil defect detection method of claim 4, wherein, When denoising the image samples, a Gaussian filtering algorithm, a median filtering algorithm or a mean filtering algorithm is used to denoise the image samples. 6.The UNet model-based printed fat oil defect detection method of claim 4, wherein, When the image samples after random noise removal are subjected to contrast enhancement, a histogram equalization algorithm or an adaptive histogram equalization algorithm is used to enhance the contrast of the image samples after random noise removal. 7.The UNet model-based printed fat oil defect detection method of claim 4, wherein, After obtaining the defect-highlighted image samples, the defect edges of the images in the defect-highlighted image samples are identified to obtain potential defect areas with preliminary contours; The potential defect areas are expanded to automatically mark the defect areas of boundary fuzzy defects and the defect areas of uneven thickness, and to automatically mark the defect types of boundary fuzzy defects and the defect types of uneven thickness according to the area morphology, edge features and thickness changes of the potential defect areas.

8. The method for detecting printing fat oil defects based on the UNet model according to claim 7, wherein: When identifying the defect edges of the images in the defect-highlighted image samples, an edge detection algorithm is used to identify the defect edges of the images in the defect-highlighted image samples; and when expanding the potential defect areas, a region growing algorithm is used to expand the potential defect areas. 9.The UNet model-based printed fat oil defect detection method of claim 7, wherein, The automatically marked defect areas of boundary fuzzy defects and the defect areas of uneven thickness are compared with actual defect areas to correct the automatically marked defect areas of boundary fuzzy defects and the defect areas of uneven thickness, and the automatically marked defect types of boundary fuzzy defects and the defect types of uneven thickness are compared with actual defect types to correct the automatically marked defect types of boundary fuzzy defects and the defect types of uneven thickness, to obtain the real labels of training data.

10. An electronic device housing, characterized by The electronic device shell is detected by using the UNet model-based printed fat oil defect detection method according to any one of claims 1-9 in a printing process.

Images