Cigarette box strip two-dimensional code printing quality evaluation method based on space-frequency domain bidirectional transformation

Abstract

The invention discloses a cigarette box strip two-dimensional code printing quality evaluation method based on space-frequency domain bidirectional transformation, and the main conception is that the method comprises the steps: respectively extracting the features of a frequency domain mode and a space domain mode from the original cigarette box strip two-dimensional code data; on one hand, frequency domain modal features are subjected to frequency domain filtering, and equivalent transformation is carried out on time domain convolution in a frequency domain product form, so that the time domain modal features can be effectively embedded into a training process of a deep network model; on the other hand, the airspace modal features are combined with a multi-head attention mechanism through central differential convolution, and deep information such as details and overall context of the airspace modal features is mined; performing feature interaction enhancement on the two modal features by means of a cross attention mechanism; and finally, fusing the frequency domain information and the space domain information, and sending the fused information to a two-dimensional code printing quality evaluation network to output a final evaluation score. According to the method, specific problems in a real cigarette packaging production line scene are combined, and rapid and efficient evaluation of the cigarette box strip two-dimensional code printing quality is effectively achieved.

Description

Technical Field

[0001] The present invention relates to the field of cigarette manufacturing, and particularly to a method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation. Background Art

[0002] In the processes of cigarette production, packaging, transportation, etc., the "carton-strip" association system based on two-dimensional code technology can well realize the rapid positioning of cigarette quality problems and the whole-process tracking of production logistics, thereby improving the informatization level of cigarette products and facilitating quality traceability. And in the whole "carton-strip" association system, the barcode reader plays a crucial role, which can quickly and accurately identify the cigarette information contained in the cigarette carton and strip two-dimensional codes.

[0003] However, in the actual process of identifying cigarette carton and strip two-dimensional codes, the printing quality is the main factor affecting the identification. For example, the two-dimensional code may have problems such as light ink, scratches, linear missing blanks, uneven thickness at the four corners for positioning, duplicate codes, and wear resulting in unclear two-dimensional codes. These defects greatly reduce the recognition rate of cigarette carton and strip two-dimensional codes and the accuracy of data reading. Therefore, it is necessary to effectively detect and evaluate the printing quality of cigarette carton and strip two-dimensional codes.

[0004] Traditional detection methods have many limitations in the process of processing cigarette carton and strip two-dimensional codes, often ignoring the performance of cigarette carton and strip two-dimensional codes in the frequency space; in addition, the complex interaction relationship between the spatial domain and the frequency domain is not fully considered, and this relationship is crucial for completely and accurately describing the quality of cigarette carton and strip two-dimensional codes; furthermore, when extracting the spatial domain information of cigarette carton and strip two-dimensional codes, most of them use simple convolutional neural network stacking, which only starts from a single angle and ignores multi-scale information, making the extracted feature information have great limitations. Generally speaking, traditional methods greatly reduce the effectiveness of the extracted information, thereby affecting the accuracy and reliability of the detection results of cigarette carton and strip two-dimensional codes. Summary of the Invention

[0005] In view of the above, the present invention aims to provide a method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation to solve the aforementioned technical problems.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation, which includes:

[0008] Performing standardization processing on the obtained original cigarette carton and strip two-dimensional codes;

[0009] Based on the standardized cigarette carton and pack two-dimensional code, extract the frequency-domain modal features and perform frequency-domain depth filtering, and perform an equivalent transformation on the time-domain convolution in the form of a frequency-domain product;

[0010] Based on the standardized cigarette carton and pack two-dimensional code, extract the spatial-domain modal features and combine them with the central difference convolution and the multi-head attention mechanism to perform information mining;

[0011] During the processing of the frequency-domain modal features and the spatial-domain modal features, use the cross-attention mechanism to enhance the interaction between the two features;

[0012] After fusing the enhanced frequency-domain modal features and the spatial-domain modal features, input them into a preset cigarette carton and pack two-dimensional code printing quality evaluation network, and output the final evaluation score.

[0013] In at least one possible implementation manner, the equivalent transformation of the time-domain convolution in the form of a frequency-domain product includes:

[0014] After reconstructing the standardized cigarette carton and pack two-dimensional code, use the fast inverse Fourier transform to transform the spatial-domain two-dimensional code features into two-dimensional code spectrum features;

[0015] After performing depth filtering on the frequency-domain image using a pre-constructed mask filtering module, inverse-transform it back to the spatial-domain dimension to form a spatial-frequency conversion architecture based on the spatial-domain-frequency-domain-spatial-domain mode.

[0016] In at least one possible implementation manner, the two-dimensional code reconstruction includes:

[0017] Pre-use a base filter to divide the frequency domain into several frequency bands;

[0018] Use the Fourier transform to convert the cigarette carton and pack two-dimensional code to the frequency domain to generate a two-dimensional code spectrum diagram; among them, the components of the low-frequency band, the middle-frequency band, and the high-frequency band are distributed at different positions in space.

[0019] In at least one possible implementation manner, the information mining specifically includes:

[0020] Based on the standardized cigarette carton and pack two-dimensional code, construct a fusion feature matrix of the reflectivity and texture information of the two-dimensional code key points;

[0021] Through the central difference convolution, obtain the edge and detail information in the fusion feature matrix to obtain the spatial-domain features;

[0022] Combine the multi-head self-attention mechanism to mine the spatial context correlation information in the spatial-domain features;

[0023] Use the residual connection to fuse the spatial context correlation information with the spatial-domain features.

[0024] In at least one possible implementation manner, the method for constructing the fusion feature matrix includes:

[0025] Performing variable illumination discrete key point calculation and reflectivity calculation on the standardized cigarette carton strip two-dimensional code to obtain local texture features and reflectivity features;

[0026] After upsampling the reflectivity features, fuse them with the local texture features.

[0027] In at least one possible implementation manner, the standardization processing includes:

[0028] Obtaining image data of continuous cigarette carton strip two-dimensional codes;

[0029] Adjusting the two-dimensional code in the image data to a preset fixed size;

[0030] Adjusting the brightness and contrast of the two-dimensional code according to the on-site light environment.

[0031] Compared with the prior art, the main design concept of the present invention is to extract features in the frequency domain mode and the spatial domain mode from the original cigarette carton strip two-dimensional code data respectively; on the one hand, passing the frequency domain mode features through a frequency domain filter to perform an equivalent transformation on the time domain convolution in the form of a frequency domain product, so that it can be effectively embedded into the training process of the deep network model; on the other hand, combining the spatial domain mode features through central difference convolution and a multi-head attention mechanism to mine deep information such as its details and overall context; then enhancing the feature interaction of the two mode features by means of a cross-attention mechanism; after fusing the frequency domain and spatial domain features, sending them into a two-dimensional code printing quality evaluation network to output the final evaluation score. The present invention effectively realizes the rapid and efficient evaluation of the printing quality of cigarette carton strip two-dimensional codes in combination with specific problems in the real cigarette packaging production line scenario. Description of the Drawings

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described below in conjunction with the drawings, where:

[0033] Figure 1 is the overall process framework diagram of the method for evaluating the printing quality of cigarette carton strip two-dimensional codes based on bidirectional transformation in the spatial and frequency domains provided by the embodiment of the present invention;

[0034] Figure 2 is the flow chart of the method for mining dark information of spatial domain mode features provided by the embodiment of the present invention;

[0035] Figure 3 is the example diagram of the two-dimensional code evaluation score provided by the embodiment of the present invention. Detailed Embodiments

[0036] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where like or similar reference numerals denote like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention.

[0037] An embodiment of a method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-temporal domain bidirectional transformation is proposed by the present invention. Specifically, as Figure 1 shown in the architecture, which includes:

[0038] Step S1: Standardize the continuously acquired original cigarette carton and strip two-dimensional codes;

[0039] The general idea of this step may refer to performing standardization operations such as scale normalization, Gamma correction, and global threshold-based binarization processing on a series of continuous original cigarette carton and strip two-dimensional code data to reduce the computational complexity, improve the robustness and recognition speed of the two-dimensional code detection algorithm, and thus adapt to the two-dimensional code printing quality problems caused by different materials and printing methods in cigarette products, ensuring the consistency and reliability of the detection results.

[0040] Specifically, as Figure 1 shown in the architecture diagram, a series of continuous cigarette carton and strip two-dimensional codes can be first acquired and standardized. Considering the diverse packaging characteristics of different brand cigarette cartons and strips in the tobacco industry, this step can adapt to the printing quality problems caused by different materials and printing methods, ensuring the consistency and reliability of the detection results.

[0041] Specifically, first, the two-dimensional code is adjusted to a fixed size to facilitate stable detection in different environments. The specific calculation process is:

[0042] ;

[0043] where, is the original size of the two-dimensional code, is the target size.

[0044] Then, by adjusting the brightness and contrast of the cigarette carton and strip two-dimensional code, the black and white areas become more distinct. This helps to maintain the recognition effect of the two-dimensional code under unstable lighting conditions. The specific calculation process is:

[0045] ;

[0046] where, is a constant, is the correction coefficient. For images in a dark environment, when it can enhance the brightness, and when When in use, it can reduce exposure in a bright environment.

[0047] Through the above standardization process, the cigarette carton and strip QR code detection algorithm can maintain a consistent input format in different environments, reduce interference caused by factors such as lighting, deformation, and noise, thereby reducing computational complexity, and improving the robustness and recognition speed of the cigarette carton and strip QR code detection algorithm.

[0048] Step S2: Based on the standardized cigarette carton and strip QR code, extract frequency-domain modal features and perform frequency-domain depth filtering to equivalently transform the time-domain convolution in the form of a frequency-domain product;

[0049] The overall idea of this step can be to extract frequency-domain modal features from the cigarette carton and strip QR code data, and equivalently transform the time-domain convolution in the form of a frequency-domain product through a differentiable frequency-domain filter, so that it can be effectively embedded in the training process of the deep network model, and improve the sensitivity of the model to defects such as stripe noise and ink diffusion through frequency-domain analysis.

[0050] Specifically, by mining and processing the frequency-domain modal feature information of the cigarette carton and strip QR code, most existing QR code detection methods are based on deep learning, and the upsampling operation will lead to abnormal frequency statistical features. Therefore, this invention introduces frequency information to mine different features of the cigarette carton and strip QR code samples, and this step improves the sensitivity of the model to common defects such as stripe noise and ink diffusion in the cigarette carton and strip through frequency-domain analysis.

[0051] For another example Figure 1 As shown, first use the Fourier transform to convert the cigarette carton and strip QR code X to the frequency domain to generate a QR code spectrogram. Among them, the low-frequency components have higher energy and are concentrated in the upper left corner of the frequency spectrum space; the middle-frequency components are distributed in a banded form in the middle; the high-frequency components have lower energy and are located in the lower right corner. In some embodiments of the present invention, N manually designed binary basis filters are used to divide the frequency domain into four frequency bands: the division of the frequency bands is based on the value range of the frequency, and the four frequency bands are: low frequency, middle frequency, high frequency, and the entire frequency spectrum. Of course, in other embodiments, it can also be modified to no less than two frequency bands, and the frequency bands can be updated to adapt to different analysis requirements and improve the flexibility of feature extraction.

[0052] In addition, in order to adaptively select the frequencies of interest outside the basis filter, N learnable filters can be added to the binary basis filter, and the above two filters are combined, which can be expressed as:

[0053] ;

[0054] Among them, is normalization, aiming to limit the learnable filter values between -1 and 1.

[0055] Specifically, after reconstructing the aforementioned normalized cigarette carton strip two-dimensional code into a high-dynamic two-dimensional code, a two-dimensional code spectrum diagram is obtained. In the depth frequency domain filtering module under the frequency domain dimension branch, it focuses on solving the problem that the cigarette carton strip two-dimensional code cannot be recognized under the performance of high-frequency noise aliasing. The previously mentioned construction of a differentiable frequency domain filter performs an equivalent transformation on the time domain convolution in the form of a frequency domain product, so that it can be effectively embedded into the training process of the deep network model. The specific implementation method is as follows:

[0056] Adopt the fast inverse Fourier transform to first convert the spatial domain two-dimensional code features into two-dimensional code spectrum features:

[0057] ;

[0058] Among them, the spatial domain input image is expressed as , and the frequency domain transformation image is expressed as , and this transformation can be abstractly expressed as: .

[0059] Secondly, design a learnable and sample size adaptive mask filtering module , and this filtering module can adopt traditional deep learning operators, so it has differentiability. After deeply filtering the frequency domain image, inverse transform it to the spatial domain dimension to form a spatial-frequency conversion architecture based on the spatial domain - frequency domain - spatial domain mode:

[0060] ;

[0061] This scheme can realize the replacement of frequency domain filtering in the way of spatial domain convolution, and its theoretical basis is as follows:

[0062] ;

[0063] In this formula, the left side of the equal sign is the constructed frequency domain product filtering principle, and the right side of the equal sign is the equivalent spatial domain convolution principle. Through equivalent design, the frequency domain features of the cigarette carton strip two-dimensional code can be obtained, and the increase in computational complexity caused by large-size convolution kernels can be effectively reduced, thereby realizing the lightweight of the model.

[0064] Step S3: Based on the normalized cigarette carton strip two-dimensional code, extract spatial domain modal features and combine them with the central difference convolution and the multi-head attention mechanism to perform information mining;

[0065] The overall idea of this step can be to extract the spatial domain modal features from the standardized cigarette carton strip two-dimensional code data, and combine the spatial domain modal features through the combination of central difference convolution and multi-head attention mechanism to mine its detailed features and overall context features, and improve the model's ability to resist geometric deformation of cigarette carton strip packaging through spatial domain enhancement.

[0066] To expand, as Figure 1 shown, conduct feature mining and processing on the spatial domain modal feature information of the cigarette carton strip two-dimensional code. This step improves the model's ability to resist geometric deformation of cigarette carton strip packaging through spatial domain enhancement. The central attention difference module under the spatial domain branch focuses on solving the problems of two-dimensional code material and texture detail differences in the recognition of cigarette carton strip two-dimensional codes.

[0067] Combined with Figure 2 this, a differential attention module based on the serialized two-dimensional code key point image features can be constructed and the following process can be executed:

[0068] Step S31: Based on the standardized cigarette carton strip two-dimensional code, construct a fusion feature matrix of the reflectivity and texture information of the two-dimensional code key points;

[0069] Step S32: Through central difference convolution, obtain the edge and detail information in the fusion feature matrix to get spatial domain features;

[0070] Step S33: Combine the multi-head self-attention mechanism to mine the spatial context correlation information in the spatial domain features;

[0071] Step S34: Use residual connection to fuse the spatial context correlation information with the spatial domain features.

[0072] Regarding the first three sub-steps, specific introductions are made here again.

[0073] First, perform variable illumination discrete key point calculation and reflectivity calculation on the standardized two-dimensional code to obtain local texture features and reflectivity features, and then fuse the reflectivity features after upsampling with the local texture features.

[0074] Next, through central difference convolution, perform feature extraction operations on the fusion feature matrix Central difference convolution (CDC) is an improved convolution operation method. A central difference operator is added to the traditional convolution to better extract the edge information and detail features of the image. It was first proposed to improve the ability of convolutional neural networks in feature representation. Especially when dealing with images with delicate edges and textures, CDC can capture more detail information. The specific operation of using central difference convolution is as follows:

[0075] ;

[0076] Among them, represents the standard convolution operation, represents the difference operation of the central difference convolution kernel. is a hyperparameter that controls the weight and is used to balance the influence between the standard convolution and the central difference.

[0077] The central difference convolution combines the standard convolution and the difference operation, and adjusts the balance between the two through the hyperparameter so that when extracting features, it can capture both the overall information and highlight the edges and details. This method enables the central difference convolution to enhance the feature representation ability while retaining the edge information. The central difference part calculates the pixel differences through the convolution kernel, and the specific calculation formula is:

[0078] ;

[0079] Among them, is the pixel position of the convolution center, represents the offset of each relative position in the convolution kernel, represents the difference between the adjacent pixel value and the central pixel, is the weight in the convolution kernel.

[0080] Through the above central difference convolution, the basic spatial domain features of the cigarette carton and strip two-dimensional code can be obtained , but these features do not contain the context space information of the two-dimensional code. Therefore, we use the multi-head self-attention mechanism to mine the spatial context information of the two-dimensional code.

[0081] The basic spatial domain features are used to mine the spatial context correlation information in the modal features through the self-attention mechanism. The basic spatial domain features are converted into queries , keys and values through the fully connected layer, and the specific calculation formula is:

[0082] ;

[0083] Then, the similarity weight matrix is calculated using and vectors, and then weighted summation is performed with the vector to obtain a new feature vector. The specific calculation formula is:

[0084] ;

[0085] Each of the aforementioned calculations can be regarded as a separate head, and the corresponding modal features are obtained by concatenating the outputs of multiple heads , and the specific calculation formula is:

[0086] ;

[0087] where are the corresponding weights, d represents the dimension, and h represents the number of heads.

[0088] Step S4: During the processing of the frequency-domain modal features and the spatial-domain modal features, use the cross-attention mechanism to enhance the interaction between the two types of features;

[0089] The overall idea of this step can be that the frequency-domain modal and spatial-domain modal features are enhanced in feature interaction by means of the cross-attention mechanism, combining the complementary information in the frequency domain and the spatial domain to improve the detection ability for complex printing defects.

[0090] Combined with Figure 1 As shown, in order to further enhance the saliency of the two-dimensional code features on cigarette cartons and strips, the present invention proposes to use the cross-attention mechanism to construct a cross-fusion response module between layers to carry out in-depth interaction on the feature responses caused by texture structure, high-frequency noise, and reflectivity materials. This step effectively combines the complementary information in the frequency domain and the spatial domain to improve the recognition ability of the model for printing defects on cigarette cartons and strips.

[0091] The cross-attention mechanism can establish more effective interaction relationships in multi-modal or multi-level information processing. By associating the features of one input sequence with another sequence, it realizes the dynamic aggregation of information. It not only enhances the linkage between different information sources but also can more accurately focus on the details related to the current task when processing the context. In response to the objective environmental differences and the user's requirements for the security level, the constructed dual-stream architecture can be dynamically split and fused to achieve the task of evaluating the printing quality of two-dimensional codes on cigarette cartons and strips under dual-stream feature fusion.

[0092] Combined with the previous example, the cross-attention mechanism for feature enhancement first connects the text with the frequency-domain features and the spatial-domain features respectively to obtain and uses them to generate the weight vector The specific calculation formula is:

[0093] ;

[0094] where are network parameters, Logistic is a non-linear activation function, and LN is a fully connected layer network. These weights can extract the relevant feature information of the two from the frequency-domain and spatial-domain features.

[0095] Then, and are fused with the weight vector to obtain the feature vectors and , the specific calculation formula is:

[0096] ;

[0097] where are network parameters, and LN is a fully connected layer network.

[0098] Finally, in order to keep the original features within an ideal range, a scale factor can be used for constraint. The frequency-domain features are concatenated and fused with the frequency-domain feature vectors for enhancement to obtain the enhanced frequency-domain features , and the specific calculation formula is:

[0099] ;

[0100] At the same time, the spatial-domain features are concatenated and fused with the spatial-domain feature vectors for enhancement to obtain the enhanced spatial-domain features , and the specific calculation formula is:

[0101] ;

[0102] where is a hyperparameter selected through cross-validation, is the L2 norm, and Dropout and LN are the dropout layer and the normalization layer respectively.

[0103] Step S5: After fusing the enhanced frequency-domain modal features and the spatial-domain modal features, input them into a preset cigarette carton and strip two-dimensional code printing quality evaluation network, and output the final evaluation score.

[0104] In actual operation, as shown in Figure 1 , each interactively enhanced modal feature is input into an evaluation network with a multi-layer perceptron (MLP) architecture to output the printing quality score of the cigarette carton and strip two-dimensional code, providing a precise quality evaluation for the cigarette carton and strip two-dimensional code. In some specific implementation manners, it may include a four-layer network structure, where the first two layers are feedforward layers, and there is a RELU activation function behind each of them. The last layer uses a Logistic regression function to output the printing quality score of the cigarette carton and strip two-dimensional code, and the specific calculation formula is:

[0105] ;

[0106] where y is a continuous value in [0, 1], indicating the quality of the printing of the cigarette carton and strip two-dimensional code. For example, a score less than 0.5 indicates that the corresponding two-dimensional code has poor printing quality, while a score greater than 0.5 indicates that the corresponding two-dimensional code has good printing quality. Regarding the actual score situation after evaluation by the above-mentioned solution of the present invention, reference can be made to Figure 3The following are five examples of scores: In a, a large area of the QR code is missing, so the recognition fails and the score is 0; in b, the QR code is very blurred, so its score is 0.2; in c, the QR code reflects light and causes linear light spots, so its score is 0.4; in d, the QR code image is clear, but due to the angle tilt, its score is 0.7; in e, the QR code has the correct angle and a clear image, so its score is 0.9.

[0107] In summary, the main design concept of the present invention is to extract the features of the frequency-domain mode and the spatial-domain mode from the original cigarette carton strip QR code data respectively. On the one hand, the frequency-domain mode features are passed through a frequency-domain filter to perform an equivalent transformation on the time-domain convolution in the form of a frequency-domain product, so that it can be effectively embedded into the training process of the deep network model. On the other hand, the spatial-domain mode features are combined with the central difference convolution and the multi-head attention mechanism to mine the deep information such as details and overall context. Then, the two-mode features are enhanced by feature interaction through the cross-attention mechanism. After fusing the frequency-domain and spatial-domain features, they are sent into the QR code printing quality evaluation network to output the final evaluation score. The present invention effectively realizes the rapid and efficient evaluation of the printing quality of cigarette carton strip QR codes by combining specific problems in the real cigarette packaging production line scenario.

[0108] In the embodiments of the present invention, if there are any expressions related to directions, they are relative concepts based on the embodiments. In addition, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects and indicates that there can be three relationships. For example, A and / or B can represent the situation of A existing alone, A and B existing simultaneously, and B existing alone. Where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0109] The structure, features, and effects of the present invention have been described in detail based on the embodiments shown in the drawings above. However, the above are only the preferred embodiments of the present invention. It should be noted that for the technical features involved in the above embodiments and their preferred methods, those skilled in the art can reasonably combine and match them into a variety of equivalent solutions without departing from and changing the design concept and technical effects of the present invention. Therefore, the present invention is not limited by the scope shown in the drawings. Any changes made according to the concept of the present invention or modified into equivalent embodiments with equivalent changes still fall within the scope covered by the description and the drawings and should be within the protection scope of the present invention.

Claims

1. A method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation, characterized in that, Including: Performing standardization processing on the obtained original cigarette carton strip two-dimensional code; Based on the standardized cigarette carton strip two-dimensional code, extracting frequency-domain modal features and performing frequency-domain depth filtering, and performing equivalent transformation on time-domain convolution in the form of frequency-domain multiplication; Based on the standardized cigarette carton strip two-dimensional code, extracting spatial-domain modal features and combining central difference convolution with a multi-head attention mechanism to perform information mining; During the processing of the frequency-domain modal features and the spatial-domain modal features, using a cross-attention mechanism to perform interactive enhancement of the two features; After fusing the enhanced frequency-domain modal features and the spatial-domain modal features, inputting them into a preset cigarette carton strip two-dimensional code printing quality evaluation network, and outputting a final evaluation score.

2. The method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation according to claim 1, wherein The equivalent transformation of time-domain convolution in the form of frequency-domain multiplication includes: After reconstructing the standardized cigarette carton strip two-dimensional code, performing a fast inverse Fourier transform to convert the spatial-domain two-dimensional code features into two-dimensional code spectrum features; After performing depth filtering on the frequency-domain image using a pre-constructed mask filtering module, inverse-transforming it back to the spatial-domain dimension to form a spatial-frequency conversion architecture based on the spatial-domain-frequency-domain-spatial-domain mode.

3. The method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation according to claim 2, wherein The two-dimensional code reconstruction includes: Previously dividing the frequency domain into several frequency bands using a basis filter; Using the Fourier transform to convert the cigarette carton strip two-dimensional code to the frequency domain to generate a two-dimensional code spectrum diagram; among them, the components of the low-frequency band, the middle-frequency band, and the high-frequency band are distributed at different positions in space.

4. The method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation according to claim 1, characterized in that The information mining specifically includes: Based on the standardized cigarette carton strip two-dimensional code, constructing a fusion feature matrix of the reflectivity and texture information of the two-dimensional code key points; Through central difference convolution, obtaining the edge and detail information in the fusion feature matrix to obtain spatial-domain features; Combining with a multi-head self-attention mechanism to mine the spatial context correlation information in the spatial-domain features; Using residual connection to fuse the spatial context correlation information with the spatial-domain features.

5. The method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation according to claim 4, wherein The construction method of the fusion feature matrix includes: Performing variable illumination discrete key point calculation and reflectivity calculation on the standardized cigarette carton strip two-dimensional code to obtain local texture features and reflectivity features; After upsampling the reflectivity features, fusing them with the local texture features.

6. The method for evaluating the printing quality of cigarette carton and strip two-dimensional codes based on spatio-frequency domain bidirectional transformation according to any one of claims 1 to 5, characterized in that, The standardization processing includes: Obtaining image data of continuous cigarette carton strip two-dimensional codes; Adjusting the two-dimensional code in the image data to a preset fixed size; Adjusting the brightness and contrast of the two-dimensional code according to the on-site light environment.

Images