An AI visual recognition method and module

Abstract

The application relates to the technical field of visual detection, in particular to an AI visual recognition method and module, wherein the method first collects a surface image of a target printed product in response to completion of a silk-screen printing action, then performs image correction on the surface image to obtain a corrected image, then inputs the corrected image into a double-channel deep learning network based on a SqueezeNet architecture for recognition to obtain a recognition result, and finally, if the recognition result meets preset defect grading standards, the target printed product is determined as a substandard product. Compared with the prior art, the method combines the correction processing technology and the double-channel deep learning network, adopts the double-channel deep learning network based on the SqueezeNet architecture for image recognition on the basis of image correction, the image recognition accuracy is higher, and the problem of insufficient accuracy in the prior art depending on the YOLO algorithm recognition is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual detection. More particularly, the present application relates to an AI visual recognition method and module. BACKGROUND

[0002] A screen printing machine, also known as a silk screen printing machine, is a mechanical device that uses a silk screen printing plate for printing and plays an important role in the printing industry. The core mechanism of the screen printing machine is the silk screen printing plate. The silk screen printing plate is composed of a silk screen, a screen frame and a photosensitive glue. After applying a layer of photosensitive glue on the silk screen, the text part of the screen is made transparent through exposure and development processes, and the non-text part of the screen is blocked. During actual printing, ink is squeezed through the screen by a squeegee, and then leaks onto the surface of the printing material through the transparent screen holes, thereby forming the desired text.

[0003] In a traditional screen printing machine, the product after silk printing is usually visually inspected by manual inspection, which is inefficient and difficult to maintain high-quality production. In order to overcome the above problems, a visual detection method, system and mobile phone shell for mobile phone shell silk printing process are disclosed in Chinese patent application CN119198773B. The visual detection system is added to the mobile phone production line to realize automatic quality control. The core visual detection algorithm of the above technology uses the YOLO algorithm, which can quickly detect and identify CCD images. However, when detecting small objects such as tiny silk printing fonts or barcodes, there are obvious precision defects.

[0004] Therefore, in the application of screen printing visual detection, the existing technology has the problem of insufficient precision. SUMMARY

[0005] To solve the above technical problem of insufficient precision, the present application discloses an AI visual recognition method and module.

[0006] In a first aspect, the present application discloses an AI visual recognition method, comprising:

[0007] In response to the completion of the silk printing action, the surface image of the target printed product is collected;

[0008] The surface image is corrected to obtain a corrected image;

[0009] The corrected image is input into a dual-channel deep learning network based on SqueezeNet architecture for recognition to obtain a recognition result;

[0010] If the recognition result meets the pre-set defect grading standard, the target printed product is determined to be unqualified.

[0011] Beneficial effects: the method of the present application first carries out correction processing before image recognition of the surface image of the target printed product, so that the image input into the double-channel deep learning network can be clear and feature obvious, thereby improving the detection accuracy. On this basis, the present application uses the high-performance and high-precision characteristics of SqueezeNet kernel, and adopts a double-channel deep learning network based on SqueezeNet architecture for image recognition, overcoming the problem of insufficient accuracy in the existing technology relying on YOLO algorithm recognition.

[0012] Preferably, the surface image is subjected to image correction to obtain a corrected image, including:

[0013] According to a preset color difference compensation algorithm, the surface image is subjected to color difference compensation to obtain a color difference compensation image;

[0014] The color difference compensation image is subjected to silk screen texture removal by a non-local mean filtering method to obtain a denoising image;

[0015] According to a preset reference vector field, the denoising image is subjected to deformation or offset correction to obtain a corrected image.

[0016] Beneficial effects: for image correction, the method of the present application sets color difference compensation, denoising and deformation or offset correction, which can make the features of the corrected image obvious, and is conducive to double-channel deep learning network recognition, obtaining more accurate recognition results.

[0017] Preferably, the color difference compensation image is subjected to silk screen texture removal by a non-local mean filtering method to obtain a denoising image, including:

[0018] Extracting local structure features of the color difference compensation image;

[0019] Searching for a neighborhood block similar to the local structure features in the color difference compensation image or a predetermined local region;

[0020] Calculating the Gaussian weighted Euclidean distance between the local structure features and the neighborhood block;

[0021] Converting the Gaussian weighted Euclidean distance into a pixel weight matrix by using an exponential function;

[0022] According to the pixel weight matrix, the pixels in the color difference compensation image or the local region are denoised to obtain a denoising image.

[0023] Preferably, according to a preset reference vector field, the denoising image is subjected to deformation or offset correction to obtain a corrected image, including:

[0024] According to a preset vector field calculation model, calculating a key point vector field of the denoising image;

[0025] Calculate the deviation value of the key point vector field and the reference vector field;

[0026] Add the pixel coordinates of the denoised image to the deviation value to obtain the corrected image.

[0027] Beneficial effects: Compared with the prior art, the method of the present application designs a set of vector field calculation model, which can efficiently calculate the key point vector field of the denoised image. Depending on the deviation value of the key point vector field and the reference vector field, the method of the present application can quickly correct the deformation or offset of the image.

[0028] Preferably, before inputting the corrected image into the dual-channel deep learning network based on the SqueezeNet architecture for recognition, the method of the present application further comprises:

[0029] Obtain a sample image;

[0030] Based on the SqueezeNet architecture, construct a pre-inference model;

[0031] Input the sample image compensated by image compensation into the pre-inference model for training, and optimize the parameters of the pre-inference model;

[0032] Expand and adjust the pre-inference model with optimized parameters to obtain a dual-channel deep learning network.

[0033] Preferably, the image compensation of the sample image at least includes one or more of denoising, brightness adjustment, contrast adjustment and inherent texture removal.

[0034] Beneficial effects: Compared with the prior art, the method of the present application obtains a dual-channel deep learning network through pre-training, which can realize the reinforcement learning of the dual-channel deep learning network on the target printed product, and further improve the recognition accuracy.

[0035] Preferably, the defect grading standard includes first-level defects, second-level defects and third-level defects; wherein the first-level defects include one or more of ink leakage, plate blocking and plate sticking; the second-level defects at least include color difference exceeding the standard; and the third-level defects at least include pattern offset exceeding the standard.

[0036] Preferably, if the target printed product is determined to be unqualified, the method of the present application further comprises:

[0037] Output an artificial re-judgment request to an artificial interactive interface.

[0038] Preferably, if the artificial re-judgment result is rejection, output a rejection signal to a rejection device of a lower machine to reject the unqualified product.

[0039] In a second aspect, the present application further discloses an AI visual recognition module, comprising a processor and a memory, and the memory stores computer program instructions, which realize the AI visual recognition method of the first aspect when executed by the processor.

[0040] The present application has the following beneficial effects:

[0041] (1) Compared with the prior art, the present application utilizes the high performance and high precision characteristics of SqueezeNet kernel, adopts a double-channel deep learning network based on SqueezeNet architecture for image recognition, and overcomes the problem of insufficient precision in the prior art which relies on YOLO algorithm recognition.

[0042] (2) Compared with the prior art, the present application method sets up color difference compensation, denoising and deformation or offset correction, and these correction processes can make the features of the corrected image obvious, which is beneficial to the double-channel deep learning network recognition and obtains more accurate recognition results.

[0043] (3) Compared with the prior art, the present application method designs a vector field calculation model, which can efficiently calculate the key point vector field of the denoised image, and depending on the deviation value of the key point vector field and the reference vector field, the present application method can quickly correct the deformation or offset of the image. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a flow chart of the AI visual recognition method in the first embodiment of the present application;

[0045] Figure 2 is an effect diagram of the recognition result being ink leakage in the first embodiment of the present application;

[0046] Figure 3 is a structural schematic diagram of the AI visual recognition module in the second embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application.

[0048] Embodiment one

[0049] As shown in Figure 1 , the present embodiment discloses an AI visual recognition method, comprising:

[0050] S10: In response to the completion of the silk screen printing action, the surface image of the target printed product is collected.

[0051] In the embodiment, the method starts to execute after the screen printing equipment is started and the initialization is completed. The image acquisition device can adopt a double high-speed global color area array camera + high-speed 4K line scanning camera to complete the detection task of different products (mainly bottle and can).

[0052] S20: performing image correction on the surface image to obtain a corrected image.

[0053] In the embodiment, the specific process of image correction can include color difference compensation, texture removal, deformation, and offset correction.

[0054] S30: inputting the corrected image into a double-channel deep learning network based on a SqueezeNet architecture to perform recognition and obtaining a recognition result.

[0055] In the embodiment, the recognition result can be a semantic understanding description (i.e., a text description) of the double-channel deep learning network, or a specific numerical value.

[0056] S40: if the recognition result meets a preset defect grading standard, determining that the target printed product is unqualified.

[0057] In the embodiment, the defect grading standard includes a first-level defect, a second-level defect, and a third-level defect. The first-level defect includes ink leakage (such as shown in FIG. 1), a plug plate, and a paste plate. The second-level defect at least includes color difference exceeding a standard. The third-level defect at least includes pattern offset exceeding a standard. Figure 2

[0058] For example, if the color difference is greater than 0.5, it is determined that the color difference exceeds the standard. If the pattern offset is greater than or equal to 0.5 mm, it is determined that the pattern offset exceeds the standard.

[0059] Specifically, when the semantic understanding description matches the text description of the first-level defect, or when the output numerical value of the second-level and third-level defects exceeds the preset standard value, it is indicated that the target printed product is unqualified.

[0060] Through the steps S10-S40, the method makes the image input into the double-channel deep learning network more clear and the features more obvious through the image correction processing, improves the detection accuracy of image recognition, and uses the characteristics of the SqueezeNet kernel to enhance the image recognition capability, thereby overcoming the problem of insufficient image recognition accuracy in the prior art.

[0061] Before the step S10, the method needs to construct the double-channel deep learning network, and the specific process is as follows.

[0062] S100: obtaining a sample image.

[0063] ​​In the embodiment, the sample image is a sample image of a target printed product with qualified printing quality, and the sample image can be multiple.

[0064] S200: Construct a pre-inference model based on the SqueezeNet architecture.

[0065] S300: Input the sample image compensated by image compensation into the pre-inference model for training, and optimize the parameters of the pre-inference model.

[0066] Before the step S300, the sample image is compensated, at least including one or more of denoising, brightness adjustment, contrast adjustment and inherent texture removal.

[0067] S400: Expand and adjust the pre-inference model with optimized parameters to obtain a dual-channel deep learning network.

[0068] More specifically, the input sample image is defined as , and the compensated sample image is .

[0069] For the denoising process, the expression is:

[0070]

[0071] In the expression, represents a Gaussian filter, and represents the parameters of the Gaussian filter.

[0072] For brightness adjustment, the expression is:

[0073]

[0074] In the expression, represents an enhancement coefficient, and represents a bias coefficient.

[0075] For contrast adjustment, the expression is:

[0076]

[0077] In the expression, represents the gray value of the sample image after contrast equalization, and represents the frequency of the original gray value of the sample image.

[0078] For removing inherent texture, the expression is:

[0079]

[0080] In the expression, h represents an image enhancement processing function. ​

[0081] In addition to compensating the sample image, the sample image can also be subjected to image feature extraction and input into the pre-inference model. The specific expression is:

[0082]

[0083] In the formula, F represents the feature of the sample image, represents a feature extraction function.

[0084] Through the above steps S100-S200, the noise and interference elements in the image are removed, and the main features of the image are extracted, providing a basis for subsequent image semantic segmentation. First, the input and output of the sample image segmentation model are determined. For the input, the original sample image and its corresponding label file are used, and the label file contains defect information of different categories. For the output, the expected output is configured to obtain feature information of different defects for subsequent processing. Secondly, in order to convert the image segmentation problem into a multi-classification problem, that is, to judge whether a pixel point belongs to a certain defect category area. In order to achieve this, the SqueezeNet deep neural network is used to complete the training process of the sample image, thereby obtaining the above-mentioned pre-inference model.

[0085] During the above step S300, the sample image subjected to the above compensation can also be subjected to further preprocessing, which is a part of the image correction training link, and specifically includes color difference compensation, filter denoising, and deformation or offset correction.

[0086] Through the above image correction training link, the model parameters of the pre-inference model can be optimized, and the detection accuracy of the subsequent substantive recognition process can be improved.

[0087] The specific process of the above step S400 is:

[0088] First, the main part of the pre-inference model is retained, and a double-channel branch is constructed, which is a segmentation network branch and a difference model branch respectively. The above main part and double-channel branch are combined to finally obtain a double-channel deep learning network.

[0089] More specifically, the segmentation network branch is used to extract the defect distribution features in the complex background of the sample image, and the difference model branch is used to capture the local texture abnormalities of the silk screen part in the sample image.

[0090] For the combination of the segmentation network branch and the backbone part of the pre-inference model, a feature map can be introduced from a certain layer (such as an intermediate layer) of the SqueezeNet backbone network to construct a segmentation network branch. The segmentation network branch can adopt a structure similar to U-Net, and through upsampling and skip connection, the low-resolution feature map is restored to the same resolution as the input image to realize pixel-level defect segmentation. For example, transposed convolution is used for upsampling, and the feature maps of the corresponding layers are spliced.

[0091] For the combination of the difference model branch and the backbone part of the pre-inference model, a feature map can also be introduced from a certain layer of the backbone network to construct a difference model branch. The difference model branch can include multiple convolution layers and pooling layers, which focus on extracting local texture features of the silk printing part. For example, a small-scale convolution kernel (such as 3x3) is used for convolution operation, and then the feature map is converted into a feature vector through global average pooling.

[0092] Further, in the step S20, the surface image is subjected to image correction to obtain a corrected image, specifically:

[0093] S21: According to a preset color difference compensation algorithm, the surface image is subjected to color difference compensation to obtain a color difference compensation image.

[0094] S22: Using a non-local mean filtering method, the color difference compensation image is subjected to silk screen texture removal to obtain a denoising image.

[0095] S23: According to a preset reference vector field, the denoising image is subjected to deformation or offset correction to obtain a corrected image.

[0096] Through the above steps S21-S23, the method of the embodiment can correct the surface image in multiple dimensions, so that the features of the corrected image are more obvious, thereby facilitating the recognition of the dual-channel deep learning network and obtaining more accurate recognition results.

[0097] The color difference compensation algorithm of the above step S21 is specifically:

[0098]

[0099] It should be explained that in the recognition process, the surface image and the previous sample image share a symbol definition to realize the reinforcement iterative learning of the model. represents the surface image after color difference compensation.

[0100] In the formula, and These represent the compensation coefficients. These two coefficients are obtained during training when the sample images are input into the pre-inference model. Specifically, during pre-training, a template image is first configured, and then the Ostu algorithm is used to obtain the foreground and background thresholds of the template image. The mean background pixel value (BVal) and the mean foreground pixel value (FVal) are calculated. The compensation coefficients between the processed sample image and the template image are then calculated based on the linear relationship between the foreground and background. and .

[0101] Furthermore, the specific process in step S22 above is as follows:

[0102] S221: Extract local structural features of the color difference compensated image.

[0103] For example, with the first Centered on a target pixel, define a kernel of size k×k. , used to describe local structural features.

[0104] S222: Search for neighborhood blocks with similar local structural features within the chromatic aberration compensated image or a predetermined local region.

[0105] For example, searching for and matching within the entire image or a local area. Similar to the first Neighboring blocks , covering the periodic repeating areas of the silkscreen texture.

[0106] S223: Calculate the Gaussian weighted Euclidean distance between the local structural features and the neighboring blocks.

[0107] Specifically, the calculation formula for step S223 above is as follows:

[0108]

[0109] In the formula, This represents the Gaussian-weighted Euclidean distance between local structural features and neighboring blocks. Indicates the side length of the neighboring block. Indicates the range of neighboring blocks patch All pixels within p Perform a summation operation. This represents the Gaussian kernel function (standard deviation is...). ), Indicates the nucleus Inner pixel p eigenvalues, Represents a neighboring block Pixels p eigenvalues.

[0110] In this embodiment, assigning a higher weight to the center pixel can enhance the accuracy of structure matching.

[0111] S224: Convert the Gaussian weighted Euclidean distance into a pixel weight matrix using an exponential function.

[0112] Specifically, the specific expression of the above step S224 is:

[0113]

[0114] wherein, represents the pixel weight matrix, exp represents the exponential function, max represents the maximum value function, represents a parameter for controlling the smoothing strength, which can be optimized by training in step S300. Generally, the parameter is positively correlated with , and usually = 10 .

[0115] S225: According to the pixel weight matrix, the pixels in the color difference compensation image or the local region are denoised to obtain a denoised image.

[0116] Specifically, the specific expression of the above step S225 is:

[0117]

[0118] wherein, represents the denoised value of the th target pixel point, represents a normalization factor, represents the center pixel value of the th neighborhood block .

[0119] In the present embodiment, .

[0120] Through the image denoising process of the above steps S221-S225, the clarity of the image can be improved, the edge features of the image can be enhanced, and the recognition accuracy of the subsequent model can be further improved.

[0121] Further, the specific process of the above step S23 is:

[0122] S231: Calculate the key point vector field of the denoised image according to the preset vector field calculation model.

[0123] wherein, the calculation formula of the key point vector field is:

[0124]

[0125] wherein, represents the key point vector field, represents a summation sign of the target pixel point, a key weight of the target pixel point, a key weight of the target pixel point, a feature vector of the target pixel point, a feature vector of the target pixel point, a reference vector value of the target pixel point, a reference vector value of the target pixel point, a smoothing coefficient (for inhibiting excessive deformation), a summation sign of the target pixel point and a pixel point in a neighborhood block, a summation sign of the target pixel point and a pixel point in a neighborhood block, a summation sign of the target pixel point and a pixel point in a neighborhood block, a displacement of the target pixel point in the x-axis direction, a displacement of the target pixel point in the x-axis direction, a displacement of the target pixel point in the x-axis direction, a displacement of the target pixel point in the x-axis direction.

[0126] S232: Calculate the deviation value of the key point vector field and the reference vector field.

[0127] S233: Add the pixel coordinates of the denoised image to the deviation value to obtain a corrected image.

[0128] The calculation formula of step S233 is:

[0129]

[0130] In the formula, denotes the horizontal coordinate of a certain pixel after correction, denotes the vertical coordinate of a certain pixel after correction, denotes the horizontal coordinate of a certain pixel before correction, denotes the vertical coordinate of a certain pixel before correction, denotes the deviation value of the horizontal coordinate of a certain pixel after correction, denotes the deviation value of the vertical coordinate of a certain pixel after correction.

[0131] Further, after the above step S40, if the target printed product is determined to be a defective product, the method of the embodiment further comprises:

[0132] S50: Output an artificial re-judgment request to an artificial interactive interface.

[0133] Through the above step S50, the method of the embodiment realizes the combined judgment of AI+expert, and can reduce the misjudgment rate.

[0134] Further, if the artificial re-judgment result is a rejection, an ejection signal is output to the lower computer ejection device to eject the defective product. The ejection signal links the lower computer ejection device to realize a real-time ejection response action below 500 ms.

[0135] Through the technical solution, the unqualified products can be efficiently removed, and the product accumulation at the output end of the production line is avoided.

[0136] According to the above technical description, the working principle of the method of the embodiment is as follows:

[0137] 1. Install a detection device after the printing position of the silk screen printer, and complete preliminary work such as signal docking;

[0138] 2. Collect about 10 good product samples, and model based on the samples;

[0139] 3. Start the automatic operation mode, detect the model capability according to the pre-built model, manually rejudge the system detection result, and learn the over-inspected products online to improve the model capability;

[0140] 4. Get a stable model, and then perform the detection task.

[0141] According to the above scheme design, the method of the embodiment at least has the following beneficial effects:

[0142] 1. The detection efficiency is improved, and the single-piece detection time is reduced from 3-5 seconds to 0.8 seconds.

[0143] 2. The labor cost is saved, and the automatic AI detection replaces the 4-person quality inspection team in 2 shifts.

[0144] 3. The detection accuracy is greatly improved, and the missed detection rate is reduced from 4.2% to 0.15%.

[0145] Embodiment two

[0146] As shown in Figure 3 , the embodiment discloses an AI visual recognition module, which includes a processor and a memory, and the memory stores computer program instructions, which realize the AI visual recognition method recorded in embodiment one when executed by the processor.

[0147] More specifically, the above-mentioned AI visual recognition module further includes a detachable highlight line light source, a double high-speed global color area array camera, a high-speed 4K line scanning camera, a quality traceability database, an online learning sample library and a man-machine interaction system.

[0148] The detachable highlight line light source is used to realize the emission of a multi-angle highlight combination strip light array.

[0149] The double high-speed global color area array camera + high-speed 4K line scanning camera is used to complete the detection task of different products (including products with curved surfaces).

[0150] The above-mentioned processor adopts an embedded AI processing unit integrated with a GPU acceleration chip.

[0151] Quality traceability database for recording defect statistics output document.

[0152] Online learning sample library for improving detection model and improving detection rate

[0153] Human-computer interaction system for quickly (within 10 minutes) completing stable product modeling.

[0154] It should be further explained that the AI vision recognition module can exist independently as a separate module, independent of the screen printing machine, and can be applied to other production lines with printing or gilding functions.

[0155] Although a number of embodiments of the present application have been shown and described herein, it will be understood by those skilled in the art that such embodiments are presented by way of example only. Numerous changes, modifications and alternatives can be suggested to one skilled in the art that are yet within the spirit and scope of this application. It is understood that in the process of practicing the present application, various alternatives to the embodiments of the application described herein can be employed.

Claims

1. An AI visual recognition method, characterized in that, include: In response to the completion of the screen printing action, a surface image of the target printed product is acquired; The surface image is then corrected to obtain a corrected image; The corrected image is input into a dual-channel deep learning network based on the SqueezeNet architecture for recognition, and the recognition result is obtained. If the identification result meets the preset defect classification standard, the target printed product is determined to be a defective product; The surface image is corrected to obtain a corrected image, including: According to the preset color difference compensation algorithm, the surface image is subjected to color difference compensation to obtain a color difference compensated image; The nonlocal mean filtering method is used to remove the screen texture from the color difference compensation image to obtain a denoised image; Based on a preset reference vector field, the denoised image is deformed or offset corrected to obtain a corrected image; The nonlocal mean filtering method is used to remove the screen texture from the color difference compensation image, resulting in a denoised image including: Extract the local structural features of the color difference compensated image; Search for neighboring blocks similar to the local structural features in the chromatic aberration compensated image or a predetermined local region; The Gaussian-weighted Euclidean distance between the local structural feature and the neighboring block is calculated. The Gaussian-weighted Euclidean distance is converted into a pixel weight matrix using an exponential function; Based on the pixel weight matrix, the pixels in the color difference compensation image or local area are denoised to obtain a denoised image. Based on a preset reference vector field, the denoised image is deformed or shifted to obtain a corrected image, including: The key point vector field of the denoised image is calculated according to the preset vector field calculation model. Calculate the deviation between the key point vector field and the reference vector field; The pixel coordinates of the denoised image are added to the deviation value to obtain the corrected image; Before inputting the corrected image into a dual-channel deep learning network based on the SqueezeNet architecture for recognition, the method further includes: Acquire sample images; Based on the SqueezeNet architecture, a pre-inference model is built. The image-compensated sample images are input into the pre-inference model for training, thereby optimizing the parameters of the pre-inference model; The pre-inference model with optimized parameters is expanded and adjusted to obtain the dual-channel deep learning network.

2. The AI ​​visual recognition method according to claim 1, characterized in that, Image compensation of the sample image includes at least one or more of the following: noise reduction, brightness adjustment, contrast adjustment, and inherent texture removal.

3. The AI ​​visual recognition method according to claim 1, characterized in that, The defect classification standard includes Level 1 defects, Level 2 defects, and Level 3 defects; wherein, Level 1 defects include one or more of ink leakage, plate clogging, and plate smearing; Level 2 defects include at least color difference exceeding the standard; and Level 3 defects include at least pattern offset exceeding the standard.

4. The AI ​​visual recognition method according to claim 1, characterized in that, If the target printed product is determined to be defective, the method further includes: Output a request for manual review to the human interaction interface.

5. The AI ​​visual recognition method according to claim 4, characterized in that, If the manual review result is rejection, a rejection signal is output to the lower-level rejection device to remove the defective product.

6. An AI visual recognition module, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the AI ​​visual recognition method according to any one of claims 1-5 is implemented.

Images