An iterative optimization image recognition method based on deep learning

Abstract

The present invention relates to the field of generalization processing of image recognition optimization models, and in particular to an image iterative optimization recognition method based on deep learning, comprising: acquiring real-time images to establish real-time image features; performing initial optimization region recognition processing according to the real-time image features to establish an image feature classification database; using the image feature classification database to establish an image optimization region segmentation model based on deep learning; obtaining image iterative optimization recognition results according to the image optimization region segmentation model, extracting common features of the image through a deep learning algorithm, and combining an internal self-loop iteration mechanism to construct an image recognition model with wide applicability and high accuracy. When processing image recognition tasks, the model can not only be better generalized to unseen data, but also provide more accurate and reliable prediction results.

Description

Technical Field

[0001] The present invention relates to the field of generalization processing of image recognition optimization models, and in particular to an image iterative optimization recognition method based on deep learning. Background Art

[0002] Traditional image recognition often uses deep learning algorithms to establish classification models, and uses a large number of training sets to ensure the stability of the model structure and the accuracy of the model output. However, the single-flow model training process requires manual labeling of abnormal outputs and re-adding them to the training set to improve the model processing capabilities. This has certain limitations on the generalization ability of the recognition model, and it is impossible to have self-iterative error correction capabilities after the solution is actually deployed. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides an image iterative optimization recognition method based on deep learning. By extracting common features of images based on deep learning algorithms, the generalization ability of the final training model is enhanced, and the internal self-loop iteration of the scheme is used to further improve the accuracy of the model output.

[0004] To achieve the above object, the present invention provides an image iterative optimization recognition method based on deep learning, which is characterized by comprising:

[0005] S1, collect real-time images and establish real-time image features;

[0006] S2, performing initial optimization region recognition processing according to the real-time image features to establish an image feature classification database;

[0007] S3, using the image feature classification database to establish an image optimization region segmentation model based on deep learning;

[0008] S4. Obtaining an image iterative optimization recognition result according to the image optimization region segmentation model.

[0009] Preferably, the collecting of real-time images and establishing real-time image features comprises:

[0010] Collect real-time images;

[0011] Acquire the image resolution, image type and image capacity of the real-time image as image data features of the real-time image;

[0012] Acquire color features, texture features, and shape features of the real-time image as image content features of the real-time image;

[0013] Using the real-time image to perform digital processing to obtain a digital matrix of the real-time image;

[0014] Using the image data features, image content features and digital matrix of the real-time image as real-time image features;

[0015] Among them, image types include binary images, grayscale images, indexed images and RGB images.

[0016] Further, performing initial optimization region recognition processing according to the real-time image features to establish an image feature classification database includes:

[0017] S2-1, establishing a data label database according to the image data features of the real-time image features;

[0018] S2-2, establishing a region label database according to the image content features of the real-time image features;

[0019] S2-3, establishing a digital label database according to the digital labels of the real-time image features;

[0020] S2-4. Utilize the data label database, the region label database and the digital label database as an image feature classification database.

[0021] Furthermore, establishing a data tag database according to the image data features of the real-time image features includes:

[0022] S2-1-1, determining whether the image type of the real-time image is a binary image, if so, using the real-time image to establish an image pixel filling matrix corresponding to the image resolution according to the image data features of the real-time image features, otherwise, executing S2-1-2;

[0023] S2-1-2, determining whether the image type of the real-time image is a grayscale image, if so, using the real-time image as a real-time grayscale image, otherwise, performing grayscale processing on the real-time image to obtain a real-time grayscale image;

[0024] S2-1-3, using the image data features of the real-time image features corresponding to the image resolution to establish an image pixel filling matrix according to the real-time grayscale image;

[0025] S2-1-4, obtaining a matrix vector of the image pixel filling matrix as a secondary image data feature label;

[0026] S2-1-5. Use the image pixel filling matrix and the secondary image data feature labels as a data label database.

[0027] Furthermore, establishing a region label database according to the image content features of the real-time image features includes:

[0028] Using the image content feature corresponding to the color feature of the real-time image feature to obtain the corresponding color ratio and establish an image color ratio;

[0029] Using the image content feature of the real-time image feature to correspond to the texture feature to obtain the corresponding segmentation area to establish the texture segmentation area;

[0030] Using the image content features of the real-time image features to correspond to the shape features, segmentation processing is performed to obtain an image target area and an image non-target area;

[0031] The image color ratio, texture segmentation area, image target area and image non-target area are used as a region label database.

[0032] Furthermore, using the image feature classification database to establish an image optimization region segmentation model based on deep learning includes:

[0033] S3-1, using the image feature classification database to establish an image optimization region recognition model based on deep learning;

[0034] S3-2. Establish an image optimization region segmentation model based on the image optimization region recognition model.

[0035] Furthermore, using the image feature classification database to establish an image optimization region recognition model based on deep learning includes:

[0036] S3-1-1, using the region label database of the image feature classification database corresponding to the image target region and the image non-target region as the first data set;

[0037] S3-1-2, using the texture segmentation area corresponding to the regional label database of the image feature classification database as the second data set;

[0038] S3-1-3, using the image color ratio corresponding to the region label database of the image feature classification database as the third data set;

[0039] S3-1-4, using the first data set as input, the texture segmentation area corresponding to the image target area of ​​the second data set and the first data set as positive output, and the texture segmentation area corresponding to the image non-target area of ​​the second data set and the first data set as negative output, and training based on deep learning to obtain an image optimization area coarse recognition model;

[0040] S3-1-5. Determine whether the color ratio corresponding to the model output of the image optimized area coarse recognition model is consistent with the image color ratio corresponding to the regional label database of the image feature classification database. If so, use the image optimized area coarse recognition model as the image optimized area recognition model. Otherwise, use the model output of the image optimized area coarse recognition model corresponding to the inconsistent image color ratio to update the first data set, and return to S3-1-4.

[0041] Furthermore, establishing an image optimization region segmentation model according to the image optimization region recognition model includes:

[0042] S3-2-1, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the positive output side region corresponding to establish the positive region edge contour;

[0043] S3-2-2, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the negative output side region corresponding to establish the negative region edge contour;

[0044] S3-2-3, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the positive region edge contour as the fourth data set;

[0045] S3-2-4, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the edge contour of the negative region as the fifth data set;

[0046] S3-2-5, using the fourth data set and the fifth data set as input, and the segmentation line segments corresponding to the fourth data set and the fifth data set as output, and performing training based on deep learning to obtain an image optimization region basic segmentation model;

[0047] S3-2-6. Determine whether the model output of the image optimization region basic segmentation model has a corresponding model in the digital label database. If so, use the image optimization region basic segmentation model as the image optimization region segmentation model. Otherwise, use the model output of the image optimization region basic segmentation model that does not have a corresponding model as the model negative training set and return to S3-2-5.

[0048] Furthermore, the image iterative optimization recognition result obtained according to the image optimization region segmentation model includes:

[0049] S4-1, using the real-time image to input the image optimization region segmentation model to obtain the model output of the image optimization region recognition model and the model output of the image optimization region segmentation model respectively;

[0050] S4-2, determining whether the model output of the image optimization region segmentation model completely corresponds to the digital label database of the image feature classification database, if so, using the model output of the image optimization region segmentation model as the image iterative optimization recognition result, otherwise, executing S4-3;

[0051] S4-3. Determine whether the model output of the image optimization region recognition model completely corresponds to the digital label database of the image feature classification database. If so, use the completely corresponding digital label database of the image feature classification database to obtain the image pixel filling matrix corresponding to the model output of the image optimization region recognition model to establish an updated data set, and return to S3-2-5.

[0052] Compared with the closest prior art, the present invention has the following beneficial effects:

[0053] Using deep learning algorithms to extract features from images can effectively enhance the generalization ability of the model. In this process, by constructing a complex neural network structure, representative common features can be identified and extracted from a large amount of image data. The extraction of these features provides a rich data basis for subsequent model training, thereby ensuring the adaptability and robustness of the model when facing new data. In order to further improve the output accuracy of the model, an internal self-loop iteration mechanism is designed, which allows the model to self-evaluate and self-optimize during the training process. Through repeated iterations, the model can gradually correct the prediction error and achieve continuous performance improvement. This self-loop iteration not only enhances the performance of the model, but also ensures the output stability of the model when facing various complex scenes. In summary, by extracting common features of images through deep learning algorithms and combining them with internal self-loop iteration mechanisms, an image recognition model with wide applicability and high accuracy is constructed. When processing image recognition tasks, this model can not only better generalize to unseen data, but also provide more accurate and reliable prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of an image iterative optimization recognition method based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0055] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Embodiment 1:

[0058] The present invention provides an image iterative optimization recognition method based on deep learning, such as Figure 1 As shown, including:

[0059] S1, collect real-time images and establish real-time image features;

[0060] S2, performing initial optimization region recognition processing according to the real-time image features to establish an image feature classification database;

[0061] S3, using the image feature classification database to establish an image optimization region segmentation model based on deep learning;

[0062] S4. Obtaining an image iterative optimization recognition result according to the image optimization region segmentation model.

[0063] S1 specifically includes:

[0064] S1-1, collect real-time images;

[0065] S1-2, acquiring the image resolution, image type and image capacity of the real-time image as image data features of the real-time image;

[0066] S1-3, acquiring color features, texture features, and shape features of the real-time image as image content features of the real-time image;

[0067] S1-4, using the real-time image to perform digital processing to obtain a digital matrix of the real-time image;

[0068] S1-5, using the image data features, image content features and digital matrix of the real-time image as real-time image features;

[0069] Among them, image types include binary images, grayscale images, indexed images and RGB images.

[0070] S2 specifically includes:

[0071] S2-1, establishing a data label database according to the image data features of the real-time image features;

[0072] S2-2, establishing a region label database according to the image content features of the real-time image features;

[0073] S2-3, establishing a digital label database according to the digital labels of the real-time image features;

[0074] S2-4. Utilize the data label database, the region label database and the digital label database as an image feature classification database.

[0075] S2-1 specifically includes:

[0076] S2-1-1, determining whether the image type of the real-time image is a binary image, if so, using the real-time image to establish an image pixel filling matrix corresponding to the image resolution according to the image data features of the real-time image features, otherwise, executing S2-1-2;

[0077] S2-1-2, determining whether the image type of the real-time image is a grayscale image, if so, using the real-time image as a real-time grayscale image, otherwise, performing grayscale processing on the real-time image to obtain a real-time grayscale image;

[0078] S2-1-3, using the image data features of the real-time image features corresponding to the image resolution to establish an image pixel filling matrix according to the real-time grayscale image;

[0079] S2-1-4, obtaining a matrix vector of the image pixel filling matrix as a secondary image data feature label;

[0080] S2-1-5. Use the image pixel filling matrix and the secondary image data feature labels as a data label database.

[0081] S2-2 specifically includes:

[0082] S2-2-1, using the image content feature of the real-time image feature to correspond to the color feature to obtain the corresponding color ratio and establish an image color ratio;

[0083] S2-2-2, using the image content feature of the real-time image feature to correspond to the texture feature to obtain the corresponding segmentation area and establish a texture segmentation area;

[0084] S2-2-3, using the image content features of the real-time image features to perform segmentation processing corresponding to the shape features to obtain an image target area and an image non-target area;

[0085] S2-2-4. Utilize the image color ratio, texture segmentation area, image target area and image non-target area as a region label database.

[0086] S3 specifically includes:

[0087] S3-1, using the image feature classification database to establish an image optimization region recognition model based on deep learning;

[0088] S3-2. Establish an image optimization region segmentation model based on the image optimization region recognition model.

[0089] S3-1 specifically includes:

[0090] S3-1-1, using the region label database of the image feature classification database corresponding to the image target region and the image non-target region as the first data set;

[0091] S3-1-2, using the texture segmentation area corresponding to the regional label database of the image feature classification database as the second data set;

[0092] S3-1-3, using the image color ratio corresponding to the region label database of the image feature classification database as the third data set;

[0093] S3-1-4, using the first data set as input, the texture segmentation area corresponding to the image target area of ​​the second data set and the first data set as positive output, and the texture segmentation area corresponding to the image non-target area of ​​the second data set and the first data set as negative output, and training based on deep learning to obtain an image optimization area coarse recognition model;

[0094] S3-1-5. Determine whether the color ratio corresponding to the model output of the image optimized area coarse recognition model is consistent with the image color ratio corresponding to the regional label database of the image feature classification database. If so, use the image optimized area coarse recognition model as the image optimized area recognition model. Otherwise, use the model output of the image optimized area coarse recognition model corresponding to the inconsistent image color ratio to update the first data set, and return to S3-1-4.

[0095] S3-2 specifically includes:

[0096] S3-2-1, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the positive output side region corresponding to establish the positive region edge contour;

[0097] S3-2-2, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the negative output side region corresponding to establish the negative region edge contour;

[0098] S3-2-3, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the positive region edge contour as the fourth data set;

[0099] S3-2-4, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the edge contour of the negative region as the fifth data set;

[0100] S3-2-5, using the fourth data set and the fifth data set as input, and the segmentation line segments corresponding to the fourth data set and the fifth data set as output, and performing training based on deep learning to obtain an image optimization region basic segmentation model;

[0101] S3-2-6. Determine whether the model output of the image optimization region basic segmentation model has a corresponding model in the digital label database. If so, use the image optimization region basic segmentation model as the image optimization region segmentation model. Otherwise, use the model output of the image optimization region basic segmentation model that does not have a corresponding model as the model negative training set and return to S3-2-5.

[0102] In the present embodiment, an image iterative optimization recognition method based on deep learning is provided. The model negative training set is not planned to set positive or negative training sets in a standardized manner in order to reduce the system operation pressure after actual application in the normal process. However, there are problems such as inaccurate model output in S3-2-6. At this time, the corresponding data is obtained to establish a negative training set to adjust the model structure.

[0103] S4 specifically includes:

[0104] S4-1, using the real-time image to input the image optimization region segmentation model to obtain the model output of the image optimization region recognition model and the model output of the image optimization region segmentation model respectively;

[0105] S4-2, determining whether the model output of the image optimization region segmentation model completely corresponds to the digital label database of the image feature classification database, if so, using the model output of the image optimization region segmentation model as the image iterative optimization recognition result, otherwise, executing S4-3;

[0106] S4-3. Determine whether the model output of the image optimization region recognition model completely corresponds to the digital label database of the image feature classification database. If so, use the completely corresponding digital label database of the image feature classification database to obtain the image pixel filling matrix corresponding to the model output of the image optimization region recognition model to establish an updated data set, and return to S3-2-5.

[0107] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0109] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0110] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An image iterative optimization recognition method based on deep learning, characterized in that: include: S1, collect real-time images and establish real-time image features; S1-1, collect real-time images; S1-2, acquiring the image resolution, image type and image capacity of the real-time image as image data features of the real-time image; S1-3, acquiring color features, texture features, and shape features of the real-time image as image content features of the real-time image; S1-4, using the real-time image to perform digital processing to obtain a digital matrix of the real-time image; S1-5, using the image data features, image content features and digital matrix of the real-time image as real-time image features; Among them, image types include binary images, grayscale images, indexed images and RGB images; S2, performing initial optimization region recognition processing according to the real-time image features to establish an image feature classification database; S2-1, establishing a data label database according to the image data features of the real-time image features; S2-1-1, determining whether the image type of the real-time image is a binary image, if so, using the real-time image to establish an image pixel filling matrix corresponding to the image resolution according to the image data features of the real-time image features, otherwise, executing S2-1-2; S2-1-2, determining whether the image type of the real-time image is a grayscale image, if so, using the real-time image as a real-time grayscale image, otherwise, performing grayscale processing on the real-time image to obtain a real-time grayscale image; S2-1-3, using the image data features of the real-time image features corresponding to the image resolution to establish an image pixel filling matrix according to the real-time grayscale image; S2-1-4, obtaining a matrix vector of the image pixel filling matrix as a secondary image data feature label; S2-1-5, using the image pixel filling matrix and the secondary image data feature labels as a data label database; S2-2, establishing a region label database according to the image content features of the real-time image features; S2-2-1, using the image content feature of the real-time image feature to correspond to the color feature to obtain the corresponding color ratio and establish an image color ratio; S2-2-2, using the image content feature of the real-time image feature to correspond to the texture feature to obtain the corresponding segmentation area and establish a texture segmentation area; S2-2-3, using the image content features of the real-time image features to perform segmentation processing corresponding to the shape features to obtain an image target area and an image non-target area; S2-2-4, using the image color ratio, texture segmentation area, image target area and image non-target area as a region label database; S2-3, establishing a digital label database according to the digital labels of the real-time image features; S2-4, using the data label database, the region label database and the digital label database as an image feature classification database; S3, using the image feature classification database to establish an image optimization region segmentation model based on deep learning; S3-1, using the image feature classification database to establish an image optimization region recognition model based on deep learning; S3-1-1, using the region label database of the image feature classification database corresponding to the image target region and the image non-target region as the first data set; S3-1-2, using the texture segmentation area corresponding to the regional label database of the image feature classification database as the second data set; S3-1-3, using the image color ratio corresponding to the region label database of the image feature classification database as the third data set; S3-1-4, using the first data set as input, the texture segmentation area corresponding to the image target area of ​​the second data set and the first data set as positive output, and the texture segmentation area corresponding to the image non-target area of ​​the second data set and the first data set as negative output, and training based on deep learning to obtain an image optimization area coarse recognition model; S3-1-5, determine whether the color ratio corresponding to the model output of the image optimization region coarse recognition model is consistent with the image color ratio corresponding to the region label database of the image feature classification database. If so, use the image optimization region coarse recognition model as the image optimization region recognition model. Otherwise, use the model output of the image optimization region coarse recognition model corresponding to the inconsistent image color ratio to update the first data set, and return to S3-1-4; S3-2, establishing an image optimization region segmentation model according to the image optimization region recognition model; S4. Obtaining an image iterative optimization recognition result according to the image optimization region segmentation model.

2. The image iterative optimization recognition method based on deep learning according to claim 1, characterized in that: Establishing an image optimization region segmentation model according to the image optimization region recognition model includes: S3-2-1, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the positive output side region corresponding to establish the positive region edge contour; S3-2-2, obtaining the model output of the image optimization region recognition model and the adjacent edge pixels of the negative output side region corresponding to establish the negative region edge contour; S3-2-3, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the positive region edge contour as the fourth data set; S3-2-4, obtaining the image pixel filling matrix corresponding to the data label database of the image feature classification database of the edge contour of the negative region as the fifth data set; S3-2-5, using the fourth data set and the fifth data set as input, and the segmentation line segments corresponding to the fourth data set and the fifth data set as output, and performing training based on deep learning to obtain an image optimization region basic segmentation model; S3-2-6. Determine whether the model output of the image optimization region basic segmentation model has a corresponding model in the digital label database. If so, use the image optimization region basic segmentation model as the image optimization region segmentation model. Otherwise, use the model output of the image optimization region basic segmentation model that does not have a corresponding model as the model negative training set and return to S3-2-5.

3. The image iterative optimization recognition method based on deep learning as claimed in claim 2, characterized in that: The image iterative optimization recognition results obtained according to the image optimization region segmentation model include: S4-1, using the real-time image to input the image optimization region segmentation model to obtain the model output of the image optimization region recognition model and the model output of the image optimization region segmentation model respectively; S4-2, determining whether the model output of the image optimization region segmentation model completely corresponds to the digital label database of the image feature classification database, if so, using the model output of the image optimization region segmentation model as the image iterative optimization recognition result, otherwise, executing S4-3; S4-3. Determine whether the model output of the image optimization region recognition model completely corresponds to the digital label database of the image feature classification database. If so, use the completely corresponding digital label database of the image feature classification database to obtain the image pixel filling matrix corresponding to the model output of the image optimization region recognition model to establish an updated data set, and return to S3-2-5.

Images