Receipt information identification method and system based on machine vision

Abstract

The invention relates to the technical field of image recognition, in particular to a receipt information recognition method and system based on machine vision. The method comprises the steps of obtaining a front image and a back image of a receipt; identifying a crease area in the back image and determining a crease area in the front image based on the crease area in the back image; and using a pre-trained prediction model, taking the pixel values of the pixel points in the crease region in the back image as input, obtaining a crease compensation value of the crease region in the front image, performing compensation processing on the crease region of the front image to obtain an enhanced image, and performing information identification by using OCR based on the enhanced image. According to the invention, the receipt detection accuracy can be improved, and the situation of receipt integration failure of a receipt integration system is reduced.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a method and system for recognizing receipt information based on machine vision. Background Technology

[0002] In modern business operations, enhancing customer experience and loyalty are core objectives. To this end, many shopping malls have launched self-service points programs based on shopping receipts. Users simply take a photo of their receipt with their mobile phone, and the system automatically recognizes the information and awards points. This digital process significantly improves user engagement and operational efficiency.

[0003] The core technology for achieving this function lies in the information recognition of receipt images. Currently, the mainstream technical approach is to first identify all the text on the receipt using Optical Character Recognition (OCR) technology, and then use models such as Graph Neural Networks (GNNs) to perform semantic relationship reasoning and extract key fields such as merchant names and transaction amounts. However, the quality of receipt images uploaded by users is often inconsistent, especially since receipts are prone to wrinkles, curling, and creases during storage. These imperfections can severely interfere with the accuracy of OCR recognition. To solve this problem, existing technologies usually preprocess the image before OCR recognition, such as using algorithms like global histogram equalization or gamma correction to enhance the image. However, these global enhancement algorithms have inherent flaws: when brightening shadow areas caused by creases, areas with normal brightness may lose text details due to pixel saturation; conversely, darkening highlight areas may make shadow areas even darker, also leading to information loss. This can cause problems such as user points failure and incorrect points calculation, which not only damages the user experience but also reduces the reliability and practicality of the entire self-service points system. Summary of the Invention

[0004] To improve the effect of image enhancement and the accuracy of subsequent text recognition on receipts, this application provides a receipt information recognition method and system based on machine vision.

[0005] Firstly, this application provides a machine vision-based method for recognizing receipt information, employing the following technical solution: A machine vision-based ticket information recognition method includes: acquiring the front and back images of the ticket; identifying crease regions in the back image and determining crease regions in the front image based on the crease regions in the back image; using a pre-trained prediction model, taking the pixel values ​​of pixels in the crease regions of the back image as input, obtaining crease compensation values ​​for the crease regions in the front image, performing compensation processing on the crease regions of the front image to obtain an enhanced image, and using OCR to recognize information based on the enhanced image; The training process of the prediction model includes: acquiring a training image set, obtaining pixels from the training image set to form a first point set and a second point set, the first point set containing multiple pixels on the front side of the crease in the pre-acquired training images, the second point set containing multiple pixels on the back side of the crease in the pre-acquired training images, and mapping the prediction model based on the pixel values ​​of the pixels in the first point set and the second point set.

[0006] By acquiring the front and back images of a receipt, this method uses the creases in the back image (where there is no text interference) to accurately predict the pixel distortion (crease compensation value) caused by the corresponding crease area in the front image, and then performs precise compensation for the crease area in the front image. This method avoids indiscriminate processing of the entire image and can target and eliminate the interference of creases on text recognition without affecting the original text information, thereby significantly improving the accuracy of subsequent OCR recognition and the reliability of the system. In this process, the monotonous background of the back image is utilized to improve the accuracy of crease area acquisition and reduce the difficulty of crease area acquisition.

[0007] Optionally, identifying crease regions in the back image and determining crease regions in the front image based on crease regions in the back image includes: performing gradient calculation on the back image; identifying pixels with gradient values ​​greater than a preset threshold as crease regions; mirroring the crease regions in the back image; and using the mirrored area of ​​the crease regions in the back image as the crease regions in the front image.

[0008] By performing gradient calculations on the back image and comparing it with a preset threshold, the crease area caused by drastic changes in brightness can be quickly and accurately identified. Then, through mirroring, the back crease position can be accurately mapped onto the front image, laying a solid foundation for subsequent accurate compensation. It has the advantages of being simple to implement and having a small computational load.

[0009] Optionally, the steps for obtaining the training image set include: obtaining front and back images of a receipt with no printed text and different types of creases to obtain the training image set.

[0010] If traditional receipts are used to obtain the training image set, the text on the receipts can affect the subsequent acquisition of crease areas. Therefore, this application uses blank receipts with different types of creases as training data, thus effectively solving the problem of difficult acquisition and annotation of training data.

[0011] Optionally, the step of mapping the prediction model based on the pixel values ​​of the pixels in the first and second point sets includes: classifying the pixels in the second point set multiple times, obtaining multiple crease categories in each classification, with the number of crease categories within a preset range; for any classification, training a linear regression model based on the multiple crease categories in the classification result, and determining the evaluation index of this classification based on the accuracy of the linear regression model corresponding to the crease category; determining the optimal number of crease categories based on the evaluation index corresponding to multiple classifications; obtaining multiple crease categories with the optimal number of crease categories, and using the linear regression model of the multiple crease categories generated in this classification as the optimal regression model, with the optimal regression model corresponding to the multiple crease categories constituting the prediction model.

[0012] This step primarily involves multiple classifications of the pixels in the second set of points. The creases are categorized into several types, with each category including a group of creases with similar shapes. For example, creases with stronger bends are grouped together, while those with weaker bends are grouped together. Different linear regression models are constructed for different crease categories, and an evaluation index reflecting the classification quality is calculated based on the accuracy of the linear model corresponding to each crease category. Finally, the optimal number of categories and the corresponding model combination are found based on the evaluation index. This method can provide differentiated and refined compensation based on the specific characteristics of the creases (such as severity).

[0013] Optionally, for any classification, the average accuracy of the linear regression models corresponding to the multiple crease categories generated by that classification is used as the evaluation index for that classification.

[0014] Optionally, the elbow method can be used to determine the inflection point of the curve formed by the evaluation indices corresponding to multiple classifications as the optimal number of crease categories.

[0015] This step introduces the elbow method to determine the optimal number of crease categories. Compared to setting based on experience or traversing all possibilities, the elbow method can find an optimal balance between improving model performance and increasing computational complexity, thus automatically determining a number of categories that ensures compensation accuracy without making the model overly complex, thereby improving the efficiency of model training.

[0016] Optionally, the step of classifying the pixels in the second point set includes: clustering the pixels in the second point set to obtain multiple crease pixel clusters; and clustering the multiple crease clusters to obtain multiple crease categories.

[0017] By clustering pixels, multiple crease pixel clusters are automatically generated, with each crease pixel cluster corresponding to one crease. Subsequently, the crease clusters are clustered again to obtain multiple crease categories.

[0018] Optionally, the steps for obtaining the accuracy of the linear regression model include: selecting multiple training image crease front and back pixels in the training image set, inputting the pixel values ​​of the back pixels into the linear regression model to obtain the crease prediction values ​​of the front pixels, and determining the accuracy of the linear regression model based on the mean square error between the multiple crease prediction values ​​of the front pixels and the actual pixel values ​​of the front pixels.

[0019] Mean squared error is a classic statistical indicator that measures the difference between predicted and true values. Using it as an accuracy criterion can objectively and accurately evaluate the model's ability to predict positive pixel values.

[0020] Optionally, the step of using a pre-trained prediction model to obtain the crease compensation value of the crease region in the front image by taking the pixel values ​​of the pixels in the crease region of the back image as input includes: for any crease region in the back image, determining the cluster to which the crease belongs based on the similarity between the crease and the crease category; inputting the pixel values ​​of the pixels in the crease region into the optimal regression model corresponding to the cluster to output the compensation value of the crease region in the front image.

[0021] When processing new receipts, the system first determines which predefined category the creases on the back of the real-time receipt belong to, and then calls the optimal regression model specific to that category to perform compensation calculations. This compensation method ensures that each type of crease receives the most suitable treatment.

[0022] Secondly, this application provides a machine vision-based ticket information recognition system, which adopts the following technical solution: The machine vision-based ticket information recognition system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the machine vision-based ticket information recognition method described above.

[0023] The aforementioned machine vision-based ticket information recognition method generates a computer program, which is stored in a memory for loading and execution by a processor. Thus, a system is created based on the memory and processor for convenient use.

[0024] This application has the following technical effects: By utilizing the crease information from the back of the image, free from text interference, a model is trained to accurately predict the pixel distortion values ​​of creases on the front. During recognition, this model performs targeted compensation only on the crease area of ​​the front image, effectively eliminating creases while perfectly preserving the details of the original text area, thus significantly improving the accuracy of subsequent OCR text recognition and the reliability of the system. Attached Figure Description

[0025] Figure 1 This is a flowchart of the receipt information recognition method based on machine vision according to an embodiment of this application.

[0026] Figure 2 This is a flowchart of step S2 of the receipt information recognition method based on machine vision in an embodiment of this application.

[0027] Figure 3 This is a flowchart of step S22 of the receipt information recognition method based on machine vision in an embodiment of this application. Detailed Implementation

[0028] This application discloses a machine vision-based method for recognizing receipt information. It utilizes image information from the back of the receipt to accurately predict and compensate for pixel distortion caused by creases in the front image. By acquiring images of both sides of the receipt, identifying and analyzing crease features in the back image, a prediction model is trained that can map back crease pixel values ​​to front crease pixel values. When processing actual receipts, this model can accurately estimate the pixel perturbation caused solely by creases in the front image based on the back creases. Then, it removes this perturbation from the front image through pixel subtraction, thereby achieving accurate crease removal without affecting the original text information, significantly improving the accuracy of subsequent OCR recognition and the reliability of the system.

[0029] Reference Figure 1 A machine vision-based ticket information recognition method includes steps S1-S3.

[0030] S1: Obtain the front and back images of the receipt, identify the crease area in the back image, and determine the crease area in the front image based on the crease area in the back image.

[0031] Acquire both the front and back images of the same ticket uploaded by the user. In one exemplary implementation, the image acquisition device (such as a smartphone camera) can provide a screen guide frame or real-time feedback to guide the user to place the ticket on a flat surface and align it for shooting, minimizing initial bending and deformation.

[0032] After acquiring the images, a series of normalization operations are performed on both the front and back images. First, a grayscale conversion is performed, converting the RGB color image into a single-channel grayscale image to simplify subsequent gradient calculations.

[0033] Since creases in an image are primarily manifested as drastic changes in pixel brightness, it is necessary to locate the crease region by calculating gradients. Specifically, the Sobel operator is used to calculate the gradients of the back image in the horizontal and vertical directions, and the gradient magnitude at each pixel is calculated.

[0034] Subsequently, the global gradient magnitude histogram of the back image is calculated, and the gradient magnitude histogram of the entire image is calculated; an appropriate gradient threshold is selected, for example, the gradient threshold can be the mean of the gradient magnitude of the entire image plus a multiple of the standard deviation; pixels greater than the threshold are regarded as the back crease region.

[0035] The front and back images of the receipt are positioned correspondingly, so there is no need to obtain the crease area of ​​the front image. Instead, the area after mirroring the pixel coordinates of the crease area in the back image is taken as the crease area of ​​the front image.

[0036] S2: Using a pre-trained prediction model, the pixel values ​​of the pixels in the crease region of the back image are used as input to obtain the crease compensation value of the crease region in the front image.

[0037] Reference Figure 2 Step S2 includes steps S21-S22.

[0038] S21: Obtain a training image set, and obtain pixels from the training image set to form a first point set and a second point set. The first point set contains multiple pixels on the front side of the crease in the pre-collected training images, and the second point set contains multiple pixels on the back side of the crease in the pre-collected training images.

[0039] Multiple creases of varying degrees and directions were artificially created on several blank receipts without any printed text. Images from both sides of these sheets were then collected to obtain training images, which together formed a training image set.

[0040] The training images in the training image set come from both sides of a blank sheet of paper. Therefore, the training images in the training image set can be divided into two categories: the front image and the back image. Since the images collected in this embodiment are from blank receipt paper, there is no need to specifically distinguish between the front and back images. Either side can be selected as the front image of the blank receipt paper, and the other side as the back image.

[0041] The back side image of the training image is identified to obtain the crease region. The front side image of the training image is also identified to obtain the crease region. The pixels of the crease region in the front image are used as the first point set, and the pixels of the crease region in the back image are used as the second point set. The steps for identifying the crease region are the same as in step S1, and will not be repeated here.

[0042] S22: Map the prediction model based on the pixel values ​​of the pixels in the first and second point sets.

[0043] Reference Figure 3 Step S22 includes steps S221-S224.

[0044] S221: Classify the pixels in the second point set multiple times, and obtain multiple crease categories each time. The number of crease categories is within a preset range.

[0045] In this step: First, cluster the pixels in the crease area of ​​the back image (which can also be understood as the pixels in the second point set) to obtain multiple crease pixel clusters.

[0046] To organize discrete crease candidate pixels into physically meaningful continuous crease lines, this embodiment employs the DBSCAN clustering algorithm. This algorithm can discover clusters of arbitrary shapes without requiring a pre-specified number of clusters. Specifically, clustering is performed based on feature vectors representing the gradient information of pixels in the crease region. These feature vectors can include the spatial location and gradient direction of the pixels. In this way, spatially continuous pixels with similar gradient directions are grouped into the same cluster, and each cluster represents a specific crease. For ease of description, this cluster is defined here as a crease pixel cluster.

[0047] Subsequently, to improve the accuracy and robustness of subsequent model training, it is necessary to denoise the clustered crease pixel clusters and remove spurious noise points caused by paper defects or uneven lighting. For any crease pixel cluster, the macroscopic features of its overall gradient characteristics are first calculated. In this embodiment, the macroscopic features of the crease pixel cluster include the global gradient magnitude.

[0048] The formula for calculating the global gradient magnitude can be expressed as: In the formula: For the first The global gradient magnitude of the pixel clusters along the crease; For clusters Number of pixels; The first pixel in the crease pixel cluster Pixels in The gradient value component in the direction; The first pixel in the crease pixel cluster Pixels in The gradient value component in the direction.

[0049] Calculate each pixel within the cluster The difference in gradient direction from the overall gradient of the cluster.

[0050] Specifically, the formula for calculating the gradient directional difference can be expressed as: In the formula, Represents pixels The difference in gradient direction between it and the pixel cluster of the crease it belongs to; Represents pixels The gradient magnitude at Components in direction; Represents pixels The gradient magnitude at Components in direction; Indicates the global gradient magnitude at Components in direction; Indicates the global gradient magnitude at Components in direction; This represents the arctangent function.

[0051] In the formula, Reflects the gradient magnitude of the pixel. The components in the direction and the global gradient magnitude are in Differences in the components in the direction; Reflects the gradient magnitude of the pixel. The components in the direction and the global gradient magnitude are in The difference in the components along the direction. Based on the difference in the two directions, the arctangent function is used to obtain the angular difference between the pixel gradient magnitude and the global gradient magnitude. A larger gradient direction difference indicates a greater difference between the pixel and the pixels in the crease pixel cluster, and therefore a higher likelihood of noise. Therefore, this embodiment sets a noise threshold to remove pixels with gradient direction differences greater than the threshold, thereby improving the accuracy and precision of subsequent processing.

[0052] Finally, considering that creases of different intensities (e.g., depth, width) have varying impacts on front-side pixel values, this embodiment performs secondary classification on all crease pixel clusters after denoising. A clustering algorithm (e.g., K-Means) is used to divide all crease pixel clusters into multiple crease categories based on their global gradient magnitude. For example, multiple crease pixel clusters can be divided into two, three, or four categories. However, determining the optimal number of crease clusters to ensure that the pixels in the back image more accurately map the pixel values ​​of the creases in the front image requires further analysis. Therefore, multiple crease clusters are clustered multiple times. In this embodiment, the clusters are clustered six times, with the corresponding number of clusters (which can also be understood as the final number of crease categories) being 1, 2, 3, 4, 5, and 6, respectively. Of course, in other embodiments, even more classifications can be performed, and more different numbers of clusters can be set.

[0053] S222: For any classification, train a linear regression model based on multiple crease categories in the classification result, and determine the evaluation index of this classification based on the accuracy of the linear regression model corresponding to the crease category.

[0054] After classification, each crease category corresponds to multiple crease pixel clusters. Each crease pixel cluster contains multiple pixels. For each crease category, a linear regression model is trained based on the pixels of the corresponding crease region within that category.

[0055] In training a linear regression model, the first step is to select training data point pairs to form a training dataset. It's important to clarify that the crease regions on both sides of the training images in the training image set correspond to each other; that is, the positions of pixels in the back image (pixels in the second point set) mirror the positions in the front image. Therefore, each pixel in the second point set has a corresponding pixel in the front image of the training images. These two corresponding pixels are used as training data point pairs. A predetermined number (which can be 100,000 based on the experience of those skilled in the art) of training data point pairs are selected from the first and second point sets to form the training dataset. Linear fitting is then performed based on the training dataset to obtain the mapping function for the crease category, and this mapping function is used as the linear regression model.

[0056] For any crease category, its corresponding mapping function can be expressed as: In the formula, This represents a mapping function used to map pixels in the second point set to pixels in the first point set; A number indicating the type of crease; Represents pixels Pixel values; Indicates crease type The intensity scaling factor; Indicates crease type The brightness bias is used for overall brightness adjustment.

[0057] The mapping function is fitted using the least squares method to determine the intensity scaling factor and brightness bias in the mapping function, thus obtaining the mapping function for each type of crease pixel cluster.

[0058] After the linear regression model is trained, its prediction accuracy needs to be verified. In this embodiment, for any linear regression model corresponding to a crease category, multiple training image pixels representing the front and back sides of the crease are selected from the training image set. The pixel values ​​of the back sides are input into the linear regression model to obtain the predicted crease values ​​of the front sides. The accuracy of the linear regression model is determined based on the mean squared error between the predicted crease values ​​of the front sides and the actual pixel values ​​of the front sides. The smaller the mean squared error, the higher the accuracy of the linear regression model. Therefore, the mean squared error is negatively correlated with the accuracy of the linear regression model. In this embodiment, the reciprocal of the mean squared error is used as the accuracy of the linear regression model.

[0059] For any given classification, the average accuracy of the linear regression models corresponding to the multiple crease categories generated by that classification is used as the evaluation index for that classification.

[0060] S223: Determine the optimal number of crease categories based on the evaluation index corresponding to multiple classifications.

[0061] Because the depth and width of creases vary, different linear regression models will be developed. As the number of crease categories increases, the classification of creases becomes more detailed. Consequently, it becomes more accurate to predict the pixel values ​​of crease regions in the front image based on the pixels in the crease region in the back image. However, as the number of crease categories increases, the computational load also increases. Therefore, it is necessary to balance computational load and accuracy.

[0062] In this embodiment, a curve is plotted based on the evaluation indices corresponding to multiple classifications, showing how the evaluation indices change with the number of categories. The inflection point of the curve formed by the evaluation indices corresponding to multiple classifications is used as the optimal number of crease categories using the elbow method. The elbow method is a conventional technique in this field and will not be described in detail here.

[0063] S224: Obtain multiple crease categories with the optimal number of crease categories, and use the linear regression model of the multiple crease categories generated by this classification as the optimal regression model. The optimal regression models corresponding to the multiple crease categories constitute the prediction model.

[0064] Once the optimal number of crease categories is determined, the linear regression model of multiple crease categories corresponding to the optimal number of crease categories is taken as the optimal regression model, and the optimal regression models corresponding to multiple crease categories constitute the prediction model.

[0065] In step S1, the front and back images of the receipt have been acquired, and the crease area of ​​the back image has been determined. For ease of understanding and description, the crease area of ​​the back image of the receipt uploaded by the customer in real time is referred to as the real-time back crease area. The pixels in the real-time back crease area are clustered to obtain multiple real-time crease clusters, each representing a linear crease.

[0066] For any crease region in the back image (real-time crease cluster), the cluster to which the crease belongs is determined based on the similarity between the crease and the crease category; the pixel values ​​of the pixels in the crease region are input into the optimal regression model corresponding to the cluster, and the compensation value of the crease region in the front image is output.

[0067] According to the description in step S221, a crease category can also be understood as a cluster, which includes multiple crease pixel clusters, each crease pixel cluster representing a crease.

[0068] For any real-time crease cluster, the similarity between the crease corresponding to the real-time crease cluster and the creases corresponding to each crease pixel cluster in the crease category can be used to determine whether the crease corresponding to the real-time crease cluster belongs to that crease category.

[0069] In this embodiment, the Euclidean distance between the cluster centers of real-time crease clusters and the cluster centers of crease categories is calculated, and the reciprocal of this Euclidean distance is used as the similarity between the real-time crease clusters and crease categories. The similarity between the real-time crease clusters and each crease category is obtained, and the crease category with the highest similarity is designated as the cluster to which the real-time crease cluster belongs. Subsequently, the pixels in the real-time crease clusters can be mapped based on the optimal linear regression model corresponding to the assigned cluster, outputting the crease compensation value for the corresponding crease region pixels in the frontal image.

[0070] S3: Compensate the crease area of ​​the front image to obtain an enhanced image, and use OCR to recognize information based on the enhanced image.

[0071] For any pixel in the crease region of a frontal image, the formula for compensating it using the crease compensation value can be expressed as: In the formula, Represents pixels Pixel values ​​after compensation processing; Represents pixels in a frontal image Original pixel values; Represents pixels The corresponding crease compensation value.

[0072] After compensating for the crease area in the frontal image, an enhanced image is obtained. Then, OCR recognition can be performed on the enhanced image, and integral calculation or information archiving can be completed based on the recognized information.

[0073] The pixel values ​​of the crease in the front image are estimated by using the pixel values ​​of the crease area in the back image, and then the crease in the front image is removed through targeted processing to improve the accuracy of subsequent OCR recognition.

[0074] This application also discloses a machine vision-based ticket information recognition system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the machine vision-based ticket information recognition method according to this application.

[0075] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0076] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for recognizing receipt information based on machine vision, characterized in that, Acquire the front and back images of the receipt; identify the crease region in the back image and determine the crease region in the front image based on the crease region in the back image; use a pre-trained prediction model, taking the pixel values ​​of the pixels in the crease region of the back image as input, to obtain the crease compensation value of the crease region in the front image. The crease area of ​​the frontal image is compensated to obtain an enhanced image, and information recognition is performed using OCR based on the enhanced image; The training process of the prediction model includes: acquiring a training image set, obtaining pixels from the training image set to form a first point set and a second point set, the first point set containing multiple pixels on the front side of the crease in the pre-acquired training images, the second point set containing multiple pixels on the back side of the crease in the pre-acquired training images, and mapping the prediction model based on the pixel values ​​of the pixels in the first point set and the second point set.

2. The method for recognizing receipt information based on machine vision according to claim 1, characterized in that, Identifying crease regions in a back image and determining crease regions in a front image based on crease regions in the back image includes: performing gradient calculation on the back image; identifying pixels with gradient values ​​greater than a preset threshold as crease regions; mirroring the crease regions in the back image; and using the mirrored area of ​​the crease regions in the back image as the crease regions in the front image.

3. The receipt information recognition method based on machine vision according to claim 1, characterized in that, The steps to obtain the training image set include: obtaining the front and back images of a small receipt with no printed text and different types of creases to obtain the training image set.

4. The method for recognizing receipt information based on machine vision according to claim 1, characterized in that, The steps for mapping the pixel values ​​of pixels in the first and second point sets to derive a prediction model include: classifying the pixels in the second point set multiple times, obtaining multiple crease categories in each classification, with the number of crease categories falling within a preset range; for any given classification, training a linear regression model based on the multiple crease categories in that classification result, and determining the evaluation index for this classification based on the accuracy of the linear regression model corresponding to the crease category; determining the optimal number of crease categories based on the evaluation index corresponding to multiple classifications; obtaining multiple crease categories with the optimal number of crease categories, and using the linear regression model of the multiple crease categories generated in this classification as the optimal regression model, with the optimal regression model corresponding to the multiple crease categories constituting the prediction model.

5. The method for recognizing receipt information based on machine vision according to claim 4, characterized in that, For any given classification, the average accuracy of the linear regression models corresponding to the multiple crease categories generated by that classification is used as the evaluation index for that classification.

6. The receipt information recognition method based on machine vision according to claim 4, characterized in that, The elbow method is used to determine the inflection point of the curve formed by the evaluation indices corresponding to multiple classifications as the optimal number of crease categories.

7. The method for recognizing receipt information based on machine vision according to claim 4, characterized in that, The steps for classifying the pixels in the second point set include: clustering the pixels in the second point set to obtain multiple crease pixel clusters; and clustering the multiple crease clusters to obtain multiple crease categories.

8. The method for recognizing receipt information based on machine vision according to claim 4, characterized in that, The steps for obtaining the accuracy of a linear regression model include: selecting multiple training image crease front and back pixels from the training image set; inputting the pixel values ​​of the back pixels into the linear regression model to obtain the predicted crease values ​​of the front pixels; and determining the accuracy of the linear regression model based on the mean square error between the multiple predicted crease values ​​of the front pixels and the actual pixel values ​​of the front pixels.

9. The method for recognizing receipt information based on machine vision according to claim 1, characterized in that, Using a pre-trained prediction model, the steps to obtain the crease compensation value of the crease region in the front image by taking the pixel values ​​of the pixels in the crease region of the back image as input include: for any crease region in the back image, determining the cluster to which the crease belongs based on the similarity between the crease and the crease category; inputting the pixel values ​​of the pixels in the crease region into the optimal regression model corresponding to the cluster to output the compensation value of the crease region in the front image.

10. A machine vision-based receipt information recognition system, characterized in that, include: The processor and memory, wherein the memory stores computer program instructions, which, when executed by the processor, implement the machine vision-based ticket information recognition method according to any one of claims 1-9.

Images