Automatic image identification and correction method based on BP neural network

Abstract

The invention discloses an automatic image recognition and correction method based on a BP neural network, and the method comprises the steps: collecting a drawing needing to be recognized and corrected through a three-axis coordinate image collection system, obtaining a drawing image, and carrying out the preprocessing of the obtained drawing image; constructing a distortion correction model based on the encoder architecture to perform distortion correction processing on the drawing image, and trimming the background of the drawing image; analyzing the obtained drawing image through an image defect detection model based on a BP neural network, and judging the defect type of the drawing image; constructing an illumination removing model based on a ResUNet generative adversarial network, and constructing a wrinkle removing model; and according to the defect type of the drawing image, selecting a de-illumination model or a de-wrinkle model to process the drawing image to obtain a clear and complete drawing image, and storing the obtained clear and complete drawing image.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an automatic image recognition and correction method based on a BP neural network. Background Technology

[0002] During image capture on mobile devices, numerous uncontrollable factors cause varying degrees of distortion in the captured images. These include camera angle, complex lighting conditions, and the paper's own curling, folding, and superposition of multiple deformations, as well as more complex deformations. Distorted paper images are unsuitable for direct automated information extraction and content analysis. The emergence of paper correction algorithms has enabled the correction of distorted and illuminated images, improving the quality of captured paper images and making them closer to realistic scanned images, thus enhancing the performance of downstream tasks such as OCR. In typical downstream task systems, such as OCR systems, image preprocessing plays a crucial role, being a critical step in the recognition system. The quality of preprocessing significantly impacts the final recognition result. The above analysis reveals that factors that may adversely affect subsequent downstream tasks during image acquisition mainly originate from two aspects: geometric interference and non-geometric interference.

[0003] The less-than-ideal relative position between the image acquisition device and the acquired paper drawing leads to geometric interference caused by factors such as perspective distortion, distortion inherent in the acquisition device itself, and deformation of the paper drawing itself. However, real-world images of drawings often exhibit a wide variety of random distortions, including perspective distortion caused by the camera's viewpoint, creases or curling caused by paper deformation, multiple random small deformations, and more complex cases of multiple deformations superimposed. These deformations severely affect the geometric quality of the image, causing varying degrees of problems for character recognition in distorted drawing images. Current mainstream methods have limited correction effects, especially for localized image distortions, which still need improvement. Furthermore, for more complex image distortions, such as multiple deformations superimposed and multiple random small creases, further improvements to the correction performance of existing methods are needed. Different acquisition environments may also cause non-geometric interferences; for example, insufficient lighting may lead to increased noise in the image, while excessive lighting may produce shadows; in addition, uneven ambient lighting may also result in unsatisfactory visual effects. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides an automatic image recognition and correction method based on BP neural network. By combining the distortion correction model, the image defect detection model based on BP neural network with the delighting model and the dewrinkling model, a clearer and more complete image is obtained.

[0005] Technical solution: To achieve the above objectives, the present invention provides an automatic image recognition and correction method based on a BP neural network, comprising the following steps:

[0006] S1. A three-axis coordinate image acquisition system is used to acquire the drawings that need to be identified and corrected, and the drawings are preprocessed.

[0007] S2. Based on the encoder architecture, a distortion correction model is constructed to perform distortion correction on the drawing image and to trim the background of the drawing image.

[0008] S3. Analyze the obtained drawing images using an image defect detection model based on a BP neural network to determine the type of defects in the drawing images;

[0009] S4. Construct a de-illumination model and a de-wrinkling model based on the generative adversarial network ResUNet;

[0010] S5. Based on the type of defect in the drawing image, select either the delighting model or the dewrinkling model to process the drawing image to obtain a clear and complete drawing image, and then store the clear and complete drawing image.

[0011] Furthermore, the three-axis coordinate image acquisition system includes a camera device that automatically controls movement along the X, Y, and Z axes to capture images of the drawing. When the camera device detects the drawing, it checks whether the drawing is completely within the captured frame. If so, it does not adjust up or down along the Z axis; otherwise, it controls the camera device to rise along the Z axis until the drawing is exactly within the captured frame. When the drawing is completely within the captured frame, it controls the camera device to move along the X and Y axes until the drawing is in the center of the captured frame. When the drawing is in the center of the captured frame, the camera device captures an image of the drawing.

[0012] Furthermore, the process of constructing a distortion correction model based on the encoder architecture to correct distortion in the drawing image includes the following steps:

[0013] S1-1. Use an encoder to extract semantic feature information to obtain control points and reference points in the drawing image, and the number of control points and reference points is the same.

[0014] S1-2. The relationship between control points and reference points is processed by interpolation method, sparse mapping is transformed into backward mapping, and the original distorted drawing image is remapped into a corrected image to obtain the distorted drawing image.

[0015] S1-3. Detect the correction effect of the distorted drawing image. If the correction effect of the drawing image does not meet the predetermined requirements, repeat S1-1 to S1-2 until the correction effect of the drawing image meets the predetermined requirements.

[0016] Furthermore, in steps S1-3, detecting whether the correction effect of the distorted drawing image meets the predetermined requirements includes the following steps:

[0017] S2-1. Use edge detection technology to extract straight lines from the twisted drawing image;

[0018] S2-2. Check whether the straight lines in the image after distortion correction are horizontal or vertical to the edge lines of the image after distortion correction. Otherwise, the correction effect does not meet the predetermined requirements. If so, proceed to the next step.

[0019] S2-3. Calculate the deviation between the tilt angle of the straight line in the image after distortion correction and the preset angle. If the calculated deviation exceeds the set deviation threshold, the correction effect does not meet the predetermined requirements; if the calculated deviation does not exceed the set deviation threshold, the correction effect meets the predetermined requirements.

[0020] Furthermore, the image defect detection model based on a BP neural network for detecting defects in drawing images includes a training part and a detection part; the training part includes the following steps:

[0021] S3-1. Obtain images of the defective drawings and perform preprocessing.

[0022] S3-2. Perform morphological processing on the preprocessed drawing image with defects;

[0023] S3-3. Input the morphologically processed defective drawing image into the BP neural network for parameter training, so that the trained parameters and the parameter types in the drawing image are mapped to obtain the trained parameters.

[0024] Furthermore, the detection section includes the following steps:

[0025] S4-1. Obtain the drawing image to be inspected and perform preprocessing;

[0026] S4-2. Perform morphological processing on the preprocessed drawing images that need to be inspected;

[0027] S4-3. Input the trained parameters obtained from the training part into the image defect detection model based on BP neural network to obtain the trained image defect detection model based on BP neural network.

[0028] S4-4. Input the morphologically processed drawing image to be detected into the image defect detection model based on BP neural network to obtain the defect type of the drawing image to be detected.

[0029] Furthermore, the defect types of the drawing image include uneven lighting defects and wrinkle defects; in step S5, when an uneven lighting defect is detected in the drawing image, an anti-lighting model is used to process the drawing image to remove light, resulting in a clear and complete drawing image; when a wrinkle defect is detected in the drawing image, an anti-wrinkle model is used to process the drawing image to remove wrinkles, resulting in a clear and complete drawing image; when both uneven lighting defects and wrinkle defects are detected in the drawing image, an anti-lighting model is first used to process the drawing image to remove light, and then an anti-wrinkle model is used to process the drawing image to remove wrinkles, resulting in a clear and complete drawing image; when no defects are detected in the drawing image, it is not necessary to use an anti-lighting model or an anti-wrinkle model to process the drawing image, resulting in a clear and complete drawing image.

[0030] Beneficial Effects: This invention provides an automatic image recognition and correction method based on a BP neural network. Employing a distortion correction model, it improves the visual presentation of drawings, making them more intuitive and accurate, which is crucial for subsequent analysis and processing. The automated drawing recognition method based on a three-axis coordinate image acquisition system can quickly process large volumes of drawings, significantly improving work efficiency and saving time and manpower. Simultaneously, the automatic recognition technology reduces errors prone to occur during manual drawing processing, such as misreading numbers or overlooking details, improving the accuracy and precision of data extraction. By effectively integrating deep learning technology with business knowledge in the engineering field, it solves the recognition challenges of high-resolution, high-density, and mutually occluded information in real-world engineering drawings. Attached Figure Description

[0031] Figure 1 This is a block diagram of an image automatic recognition and correction method based on a BP neural network;

[0032] Figure 2 This is a schematic diagram of the X-axis, Y-axis, and Z-axis in a three-axis coordinate image acquisition system;

[0033] Figure 3 This is a schematic diagram of the image defect detection model based on a BP neural network.

[0034] Figure 4 This is a schematic diagram of a BP neural network structure;

[0035] Figure 5 This is a schematic diagram of a neuron structure. Detailed Implementation

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] like Figure 1 As shown, an automatic image recognition and correction method based on a BP neural network includes the following steps:

[0038] S1. A three-axis coordinate image acquisition system is used to acquire the drawings that need to be identified and corrected, and the drawings are preprocessed.

[0039] S2. Based on the encoder architecture, a distortion correction model is constructed to perform distortion correction on the drawing image obtained in S1, and the background of the drawing image is trimmed to retain only the drawing image itself.

[0040] S3. Analyze the drawing image obtained in S2 using an image defect detection model based on a BP neural network to determine the type of defect in the drawing image;

[0041] S4. Construct a de-illumination model based on a generative adversarial network (GAN) using ResUNet++, and simultaneously construct a de-wrinkling model. The de-wrinkling model can be based on deformation field repair or generated by GAN, which is a deep learning-based algorithm model.

[0042] S5. Based on the type of defect in the drawing image, select either the delighting model or the dewrinkling model to process the drawing image obtained in S3, so as to obtain a clear and complete drawing image. Store the clear and complete drawing image to facilitate the subsequent recognition of characters in the drawing image.

[0043] In step S3, after the image defect type of the drawing image is detected by the BP neural network-based image defect detection model, the processes of removing illumination and removing wrinkles can be performed sequentially. When the defect type of the drawing image is a uniform illumination defect, the weight of illumination removal in the illumination removal model can be increased and the weight of wrinkle removal in the wrinkle removal model can be decreased during illumination removal. Conversely, when the defect type of the drawing image is a wrinkled defect, the weight of illumination removal in the illumination removal model can be decreased and the weight of wrinkle removal in the wrinkle removal model can be increased during illumination removal.

[0044] like Figure 2As shown, the three-axis coordinate image acquisition system includes a camera device that automatically controls movement along the X, Y, and Z axes to capture images of the drawing. When the camera device detects the drawing, it checks whether the drawing is completely within the captured frame. If so, it does not adjust up or down along the Z-axis; otherwise, it controls the camera device to rise along the Z-axis until the drawing is exactly within the captured frame, i.e., moving along the positive direction of the Z-axis. When the drawing is completely within the captured frame, the camera device is controlled to move along the X and Y axes until the drawing is in the center of the captured frame. When the drawing is in the center of the captured frame, the camera device captures an image of the drawing. Placing the drawing in the center of the captured frame reduces distortion and blur, improving image quality; ensures focus sharpness, simplifying subsequent processing operations; and maximizes the uniformity of illumination, reducing the computational load of subsequent illumination removal operations.

[0045] When the drawing is completely within the captured frame, the distance between the top edge of the drawing and the top edge of the captured frame is detected to obtain the top edge difference; the distance between the bottom edge of the drawing and the bottom edge of the captured frame is detected to obtain the bottom edge difference; by comparing the top edge difference and the bottom edge difference, the camera device is adjusted to move along the Y-axis.

[0046] When the difference between the upper and lower edges is greater than the difference between the lower and upper edges, the camera is moved one unit distance along the positive Y-axis; then the difference between the upper and lower edges is compared again until the difference between the upper and lower edges equals the difference between the lower and upper edges. When the difference between the upper and lower edges is less than the difference between the lower and upper edges, the camera is moved one unit distance along the negative Y-axis; then the difference between the upper and lower edges is compared again until the difference between the upper and lower edges equals the difference between the lower and upper edges.

[0047] Similarly, the distance between the left edge of the drawing and the left edge of the captured image is measured to obtain the left edge difference; the distance between the right edge of the drawing and the right edge of the captured image is measured to obtain the right edge difference; by comparing the left edge difference and the right edge difference, the camera device is moved along the X-axis.

[0048] When the difference between the left and right edges is greater than the difference between the right and left edges, the camera is moved one unit distance in the negative X-axis direction; then the differences between the left and right edges are compared again until the difference between the left and right edges equals the difference between the right and left edges. When the difference between the left and right edges is less than the difference between the right and left edges, the camera is moved one unit distance in the positive X-axis direction; then the differences between the left and right edges are compared again until the difference between the left and right edges equals the difference between the right and left edges. This process adjusts the drawing to the center of the captured image.

[0049] The flexible angle adjustment, comprehensive image capture capabilities, and easier-to-control lighting conditions of a three-axis coordinate image acquisition system make it an ideal choice for high-quality acquisition of industrial drawings, enabling better image acquisition, geometric information gathering, and lighting correction. First, configure and prepare the three-axis coordinate image acquisition system, ensuring its system calibration, measurement accuracy, and stability. In the acquisition environment, control lighting and clear away debris to ensure stable operation of the image acquisition system. Prepare the drawings, selecting appropriate sizes and types, ensuring they are flat and unfolded, and design a fixing device to maintain stability during the acquisition process. Set shooting parameters, including exposure time, aperture, and resolution, to obtain clear, uniformly bright images. Acquire drawing information, including image and geometric information, through the image acquisition system, performing preprocessing such as noise reduction and enhancement, and lighting correction to improve image consistency. Finally, optimize and evaluate the acquired data using algorithms to verify its accuracy and robustness in actual acquisition. The efficient, accurate, and stable acquisition of drawing information through the drawing information acquisition system provides reliable data support for subsequent image processing and analysis.

[0050] A three-axis coordinate image acquisition system can use a three-axis high-speed document scanner. The three-axis high-speed document scanner has an electric lifting function, which can adjust the height by selecting different paper sizes, so that the industrial-grade high-definition camera can capture clear images for subsequent light removal and wrinkle removal processing.

[0051] The process of constructing a distortion correction model based on the encoder architecture to correct distortion in drawing images includes the following steps:

[0052] S1-1. Use an encoder to extract semantic feature information to obtain control points and reference points in the drawing image. The number of control points and reference points is the same. The control points and reference points describe the shape of the drawing before and after correction, respectively.

[0053] S1-2. The relationship between control points and reference points is processed by interpolation method, sparse mapping is transformed into backward mapping, and the original distorted drawing image is remapped into a corrected image to obtain the distorted drawing image.

[0054] S1-3. Detect the correction effect of the distorted drawing image. If the correction effect of the drawing image does not meet the predetermined requirements, repeat S1-1 to S1-2 until the correction effect of the drawing image meets the predetermined requirements.

[0055] The control points are flexible and controllable, facilitating human interaction for adjustments. The cyclic twisting correction operation enhances practicality, thus mitigating the operational limitations of end-to-end methods. Post-processing methods and vertex counts can be flexibly selected based on the needs of different application scenarios. Compared to pixel-by-pixel regression, control point regression is more practical and efficient.

[0056] In steps S1-3, detecting whether the correction effect of the distorted drawing image meets the predetermined requirements includes the following steps:

[0057] S2-1. Use edge detection technology or Hough transform to extract straight lines from the distortion-corrected drawing image;

[0058] S2-2. Check whether the straight lines in the image after distortion correction are horizontal or vertical to the edge lines of the image after distortion correction. Otherwise, the correction effect does not meet the predetermined requirements. If so, proceed to the next step.

[0059] S2-3. Calculate the deviation between the tilt angle of the straight line in the corrected drawing and the preset angle. If the calculated deviation exceeds the set deviation threshold, the correction effect does not meet the predetermined requirements; if the calculated deviation does not exceed the set deviation threshold, the correction effect meets the predetermined requirements. The deviation threshold can be set according to the actual situation, and can be set to a deviation of 0.5 degrees.

[0060] like Figure 3 As shown, the image defect detection model based on a BP neural network for detecting defects in drawing images includes a training part and a detection part; the training part includes the following steps:

[0061] S3-1. Obtain images of the defective drawings and perform preprocessing.

[0062] S3-2. Perform morphological processing on the preprocessed drawing image with defects;

[0063] S3-3. Input the morphologically processed defective drawing image into the BP neural network for parameter training, so that the trained parameters and the parameter types in the drawing image are mapped to obtain the trained parameters.

[0064] The detection process includes the following steps:

[0065] S4-1. Obtain the drawing image to be inspected and perform preprocessing;

[0066] S4-2. Perform morphological processing on the preprocessed drawing images that need to be inspected;

[0067] S4-3. Input the trained parameters obtained from the training part into the image defect detection model based on BP neural network to obtain the trained image defect detection model based on BP neural network.

[0068] S4-4. Input the morphologically processed drawing image to be detected into the image defect detection model based on BP neural network to obtain the defect type of the drawing image to be detected.

[0069] like Figure 4 As shown, a BP neural network is a process of information propagation in both forward and backward directions. It typically consists of three layers: an input layer, a hidden layer, and an output layer. When the input layer receives information from the outside, the neural network begins the training process of classification and judgment, training with multiple input samples. When the actual output does not meet the expected output, it will enter the backpropagation stage of error. The error passes through the output layer and corrects the weights of each layer in a way that reduces the error. This process is repeated layer by layer in the hidden and input layers. The weights of each layer are continuously adjusted through forward information propagation and backward error propagation, which is also the process of neural network learning and training. The end of this process is marked when the network output error is reduced to an acceptable level or when the preset number of learning iterations is reached.

[0070] like Figure 5 As shown, the BP neural network is a typical multilayer feedforward network operation model trained using the backpropagation algorithm. Based on learning the gradient descent rule, it uses backpropagation to continuously adjust the lexical values ​​and weights, minimizing the sum of squared errors of the neural network. The BP neural network consists of a large number of interconnected neurons. In each neuron, x1~xi are the components of the input vector, w1~wi are the weights of each synapse, b is the bias (also called the threshold), f(Sj) is the transfer function, and yi represents the output of each neuron. The output value of each neuron node is related to the output value of the upper-layer nodes, the connection weights between two neuron nodes, the threshold or bias of the current node, and the transfer function. The overall representation is shown below:

[0071]

[0072]

[0073] In the formula, n is the number of neurons.

[0074] The defect types of the drawing image include uneven lighting defects and wrinkle defects. In step S5, when an uneven lighting defect is detected, a delighting model is used to process the drawing image to remove the lighting, resulting in a clear and complete drawing image. When a wrinkle defect is detected, a wrinkle removal model is used to process the drawing image to remove the wrinkles, resulting in a clear and complete drawing image. If only one defect is detected in the drawing image, only this one defect is processed, and no further defect processing is performed, avoiding unnecessary defect processing operations and improving the processing efficiency of the drawing image. However, excessive delighting and wrinkle removal operations can also cause new defects in the drawing image, making the text and images on the drawing image unclear.

[0075] When the defects detected in the drawing image are uneven lighting defects and wrinkle defects, the drawing image is first processed by the anti-lighting model, and then processed by the anti-wrinkle model to obtain a clear and complete drawing image. When no defects are detected in the drawing image, it is not necessary to use the anti-lighting model and the anti-wrinkle model to obtain a clear and complete drawing image.

[0076] In the process of removing illumination, the aforementioned illumination removal model can perform layout analysis on the drawing image, dividing the entire drawing image into image areas, text areas, and background areas. Illumination removal is then performed sequentially on the image areas, text areas, and background areas, as illumination factors have a greater impact on image areas, followed by text areas, and the least impact on background areas. Therefore, different degrees of illumination removal operations are required depending on the degree of interference from illumination factors in different areas.

[0077] First, a distortion correction model is used to process the drawing image, eliminating geometric distortions and improving feature localization accuracy, laying a geometric foundation for subsequent processing. Then, the defect types in the drawing image are detected, and the image is de-illuminated based on the defect type. De-illumination eliminates the interference of lighting on the drawing image, highlighting the texture and avoiding interference from uneven lighting in the wrinkle removal operation. The visual appearance of wrinkles is significantly affected by lighting; the same wrinkle may exhibit completely different grayscale distributions under different lighting conditions. After de-illumination, the image grayscale distribution is more concentrated on the reflective properties of the object itself, allowing the wrinkle removal model to focus more on the morphological features of the wrinkles rather than artifacts caused by lighting. Then, the wrinkle removal model is used for wrinkle removal processing. Wrinkles may cause local image distortion or content occlusion. If geometric deformation is not corrected first, wrinkle removal may fail due to coordinate misalignment. After correction, the geometric structure of the image is clear, and the wrinkle removal model can fill the wrinkled areas in the correct direction, avoiding content distortion. Meanwhile, geometric accuracy is improved by sequentially using a distortion correction model, a delighting model, and a dewrinkling model, thereby enhancing feature recognizability and reducing computational complexity.

[0078] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. An automatic image recognition and correction method based on a BP neural network, characterized in that: Includes the following steps: S1. A three-axis coordinate image acquisition system is used to acquire the drawings that need to be identified and corrected, and the drawings are preprocessed. S2. Based on the encoder architecture, a distortion correction model is constructed to perform distortion correction on the drawing image and to trim the background of the drawing image. S3. Analyze the obtained drawing images using an image defect detection model based on a BP neural network to determine the type of defects in the drawing images; S4. Construct a de-illumination model and a de-wrinkling model based on the generative adversarial network ResUNet; S5. Based on the type of defect in the drawing image, select either the delighting model or the dewrinkling model to process the drawing image to obtain a clear and complete drawing image, and then store the clear and complete drawing image.

2. The image automatic recognition and correction method based on BP neural network according to claim 1, characterized in that: The three-axis coordinate image acquisition system includes a camera device that automatically controls movement along the X, Y, and Z axes to capture images of the drawing. When the camera device detects the drawing, it checks whether the drawing is completely within the captured frame. If so, it does not adjust up or down along the Z axis; otherwise, it controls the camera device to rise along the Z axis until the drawing is exactly within the captured frame. When the drawing is completely within the captured frame, it controls the camera device to move along the X and Y axes until the drawing is in the center of the captured frame. When the drawing is in the center of the captured frame, the camera device captures an image of the drawing.

3. The image automatic recognition and correction method based on BP neural network according to claim 1, characterized in that: The process of constructing a distortion correction model based on the encoder architecture to correct distortion in drawing images includes the following steps: S1-1. Use an encoder to extract semantic feature information to obtain control points and reference points in the drawing image, and the number of control points and reference points is the same. S1-2. The relationship between control points and reference points is processed by interpolation method, sparse mapping is transformed into backward mapping, and the original distorted drawing image is remapped into a corrected image to obtain the distorted drawing image. S1-3. Detect the correction effect of the distorted drawing image. If the correction effect of the drawing image does not meet the predetermined requirements, repeat S1-1 to S1-2 until the correction effect of the drawing image meets the predetermined requirements.

4. The image automatic recognition and correction method based on BP neural network according to claim 3, characterized in that: In steps S1-3, detecting whether the correction effect of the distorted drawing image meets the predetermined requirements includes the following steps: S2-1. Use edge detection technology to extract straight lines from the twisted drawing image; S2-2. Check whether the straight lines in the image after distortion correction are horizontal or vertical to the edge lines of the image after distortion correction. Otherwise, the correction effect does not meet the predetermined requirements. If so, proceed to the next step. S2-3. Calculate the deviation between the tilt angle of the straight line in the image after distortion correction and the preset angle. If the calculated deviation exceeds the set deviation threshold, the correction effect does not meet the predetermined requirements; if the calculated deviation does not exceed the set deviation threshold, the correction effect meets the predetermined requirements.

5. The image automatic recognition and correction method based on BP neural network according to claim 1, characterized in that: The image defect detection model based on a BP neural network for detecting defects in drawing images includes a training part and a detection part; the training part includes the following steps: S3-1. Obtain images of the defective drawings and perform preprocessing. S3-2. Perform morphological processing on the preprocessed drawing image with defects; S3-3. Input the morphologically processed defective drawing image into the BP neural network for parameter training, so that the trained parameters and the parameter types in the drawing image are mapped to obtain the trained parameters.

6. The image automatic recognition and correction method based on BP neural network according to claim 5, characterized in that: The detection process includes the following steps: S4-1. Obtain the drawing image to be inspected and perform preprocessing; S4-2. Perform morphological processing on the preprocessed drawing images that need to be inspected; S4-3. Input the trained parameters obtained from the training part into the image defect detection model based on BP neural network to obtain the trained image defect detection model based on BP neural network. S4-4. Input the morphologically processed drawing image to be detected into the image defect detection model based on BP neural network to obtain the defect type of the drawing image to be detected.

7. The image automatic recognition and correction method based on BP neural network according to claim 1, characterized in that: The defect types of the drawing image include uneven lighting defects and wrinkle defects. In step S5, when an uneven lighting defect is detected, a delighting model is used to process the drawing image to remove the lighting, resulting in a clear and complete drawing image. When a wrinkle defect is detected, a wrinkle removal model is used to process the drawing image to remove the wrinkles, resulting in a clear and complete drawing image. When both uneven lighting defects and wrinkle defects are detected, the drawing image is first processed using a delighting model, and then processed using a wrinkle removal model to remove the wrinkles, resulting in a clear and complete drawing image. When no defects are detected in the drawing image, no processing using a delighting model or a wrinkle removal model is required, resulting in a clear and complete drawing image.

Images