A close-up image recognition method and system based on deep learning
Through the close-up graph recognition method based on deep learning, the deep convolutional neural network is used to identify close-up graphs, which solves the problems of low accuracy and limitations of traditional methods, and realizes efficient and accurate close-up graph recognition, which is suitable for different numbers of pictures Scene.
Patent Information
- Application Number
- CN202111623507.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-28
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2041-12-28
AI Technical Summary
Traditional image processing algorithms have low accuracy and strong limitations when identifying close-up images, and cannot effectively deal with the uncertain number of close-up images, and the error recognition rate is high, resulting in the intelligent audit algorithm obtaining incorrect results.
Using a close-up graph recognition method based on deep learning, the puzzles to be distinguished are split into original sub-graph groups, and the sub-graphs are input into the trained classification model for judgment, and the features are extracted using deep convolutional neural network to perform pixels Level classification, determine whether each sub-picture is a close-up picture.
It realizes the rapid and effective identification of close-up images in the illegal evidence link, which is highly robust, can handle scenes of close-up images in different numbers, has fast calculation speed and excellent results.
Smart Images
Figure CN114511730B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing and deep learning, and in particular to a close-up image recognition method and system based on deep learning. Background Art
[0002] In practical applications such as the review of traffic violation evidence images, different devices in different places and different violation categories set different rules for making violation evidence puzzles. Most of the devices will provide 1 to 5 close-up evidence sub-images in the puzzle (a small number of them even have no close-up images), and the remaining sub-images form a stream of violation evidence images arranged in chronological order. Close-up images can help find illegal vehicles and identify the license plate numbers of illegal vehicles, while the remaining sub-images provide functions such as judging the logic of the violation process and supplementing the identification of license plates. Obviously, close-up images that are mistakenly inserted into the time-sequential sub-image evidence chain, or non-close-up images that are mistakenly identified as close-up images, will cause the intelligent review algorithm to obtain incorrect results. Therefore, it is necessary to accurately distinguish between close-up images and non-close-up images in the original sub-images after the puzzle is split.
[0003] For such problems, traditional methods generally use traditional image processing algorithms, such as calculating image features such as HOG, HARR, etc. to determine similarity, using SIFT features to match key points, or even statistically analyzing the distribution of close-up images to give empirical locations.
[0004] However, traditional methods have low accuracy and strong limitations, and require the adjustment of a large number of empirical parameters. All of them require strong assumptions such as a fixed number of close-up images and the existence of sub-images that can match close-up images. For example, the features of the feature method are not very effective in many scenarios. The SIFT method cannot process close-up images that are not from local magnifications of specific sub-images. All traditional methods cannot handle situations where the number of close-up images is uncertain, such as no close-up images, or multiple close-up images. Summary of the invention
[0005] In view of this, the present invention provides a close-up image recognition method and system based on deep learning, which can quickly and effectively identify close-up images in the chain of illegal evidence with high robustness and good effect.
[0006] The present invention discloses a close-up image recognition method based on deep learning, comprising:
[0007] Split the puzzle to be distinguished into groups of original sub-pictures to be distinguished;
[0008] The original sub-image group is input into the trained classification model to determine whether each original sub-image is a close-up image.
[0009] Preferably, the training step of the classification model includes:
[0010] Collect the original puzzle picture;
[0011] According to the jigsaw original image, each sub-image in each group of original sub-images is divided into different categories;
[0012] According to each sub-image in each group of original sub-images, a trained classification model is obtained.
[0013] Preferably, collecting the original puzzle image includes:
[0014] Collect original puzzle images of traffic violation evidence chain during the day and at night in different puzzle and close-up modes from various devices.
[0015] Preferably, the method of dividing each sub-image in each group of original sub-images into different categories according to the jigsaw original image includes:
[0016] Splitting each of the jigsaw original images into a set of original sub-images;
[0017] All original sub-images in each group of original sub-images are divided into different categories through manual annotation.
[0018] Preferably, all original sub-graphs in each group of original sub-graphs are divided into different categories by manual annotation, including:
[0019] If the original sub-image is a close-up image, its corresponding category is marked as 0;
[0020] If the original sub-image is a non-close-up image, its corresponding category is marked as 1.
[0021] Preferably, obtaining a trained classification model according to each sub-image in each group of original sub-images includes:
[0022] When the number of original sub-images in each group of original sub-images is less than 8, fill them with all-black fake sub-images;
[0023] Each sub-image in each group of original sub-images is scaled to a fixed size, and then each scaled sub-image is horizontally spliced along its length to obtain a horizontal spliced image; the category corresponding to the completely black fake sub-image is marked as -1;
[0024] According to the category corresponding to each sub-image in the horizontal spliced image, a trained classification model is obtained.
[0025] Preferably, obtaining a trained classification model according to the category corresponding to each sub-image in the horizontal spliced image includes:
[0026] From left to right, set a label of the horizontal mosaic image according to the category corresponding to each sub-image in the horizontal mosaic image;
[0027] Input the horizontal splicing image and its corresponding label into a deep convolutional neural network to extract features and obtain a 1*8 feature map;
[0028] Use the obtained 1*8 feature map to train the pixel-level classification model to obtain a trained classification model.
[0029] Preferably, the label corresponding to the all-black fake sub-image is -1; when training the classification model, the -1 label is ignored.
[0030] Preferably, the determining whether each original sub-image is a close-up image includes:
[0031] If the category corresponding to the original sub-image is 0, the original sub-image is judged as a close-up image;
[0032] If the category corresponding to the original sub-image is 1, the original sub-image is judged as a non-close-up image.
[0033] The present invention also discloses a close-up image recognition system based on deep learning, comprising:
[0034] A splitting module, used for splitting the puzzle to be distinguished into original sub-picture groups to be distinguished;
[0035] The judgment module is used to input the original sub-image group into the trained classification model to judge whether each original sub-image is a close-up image.
[0036] Due to the adoption of the above technical scheme, the present invention has the following advantages: it can quickly and effectively identify close-up images in the chain of illegal evidence, has high robustness and good effect, the model calculation speed based on GPU or NPU is fast, and can smoothly process scenes with different numbers of close-up images. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0038] Figure 1 A schematic diagram of a flow chart of a close-up image recognition method based on deep learning according to an embodiment of the present invention;
[0039] Figure 2 Schematic diagram of the structure of a close-up image recognition system based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION
[0040] The present invention is further described in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.
[0041] See also Figure 1 The present invention provides an embodiment of a close-up image recognition method based on deep learning. Specifically, the method includes:
[0042] S101: splitting the puzzle to be distinguished into original sub-picture groups to be distinguished;
[0043] S102: Input the original sub-image group into the trained classification model to determine whether each original sub-image is a close-up image.
[0044] In this embodiment, the training steps of the classification model include:
[0045] Collect the original puzzle picture;
[0046] According to the original puzzle image, each sub-image in each group of original sub-images is divided into different categories;
[0047] According to each sub-image in each group of original sub-images, a trained classification model is obtained.
[0048] In this embodiment, collecting the original puzzle image includes:
[0049] Collect original puzzle images of traffic violation evidence chain during the day and at night in different puzzle and close-up modes from various devices.
[0050] In this embodiment, according to the original puzzle image, each sub-image in each group of original sub-images is divided into different categories, including:
[0051] Split each puzzle original image into a set of original sub-images;
[0052] All original sub-images in each group of original sub-images are divided into different categories through manual annotation.
[0053] In this embodiment, all original sub-images in each group of original sub-images are divided into different categories through manual annotation, including:
[0054] If the original sub-image is a close-up image, its corresponding category is marked as 0;
[0055] If the original sub-image is a non-close-up image, its corresponding category is marked as 1.
[0056] In this embodiment, a trained classification model is obtained according to each sub-image in each group of original sub-images, including:
[0057] When the number of original sub-images in each group of original sub-images is less than 8, fill them with all-black fake sub-images;
[0058] Each sub-image in each group of original sub-images is scaled to a fixed size, and then each scaled sub-image is horizontally spliced along its length to obtain a horizontal spliced image; the category corresponding to the completely black fake sub-image is marked as -1;
[0059] According to the category corresponding to each sub-image in the horizontal splicing image, a trained classification model is obtained.
[0060] In this embodiment, a trained classification model is obtained according to the category corresponding to each sub-image in the horizontal spliced image, including:
[0061] From left to right, set the label of the horizontal mosaic image according to the category corresponding to each sub-image in the horizontal mosaic image;
[0062] The horizontal splicing image and its corresponding labels are input into the deep convolutional neural network to extract features and obtain a 1*8 feature map;
[0063] Use the obtained 1*8 feature map to train the pixel-level classification model to obtain a trained classification model.
[0064] In this embodiment, the label corresponding to the completely black fake sub-image is -1; when training the classification model, the -1 label is ignored.
[0065] In this embodiment, determining whether each original sub-image is a close-up image includes:
[0066] If the category corresponding to the original sub-image is 0, the original sub-image is judged as a close-up image;
[0067] If the category corresponding to the original sub-image is 1, the original sub-image is judged as a non-close-up image.
[0068] The present invention also discloses a close-up image recognition system based on deep learning, comprising:
[0069] A splitting module 201 is used to split the puzzle to be distinguished into original sub-picture groups to be distinguished;
[0070] The judgment module 202 is used to input the original sub-image group into the trained classification model to judge whether each original sub-image is a close-up image.
[0071] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the computer-readable recording medium includes any mechanism for storing or transmitting information in a form readable by a computer (e.g., a computer).
[0072] 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. A close-up image recognition method based on deep learning, characterized in that: include: Split the puzzle to be distinguished into groups of original sub-pictures to be distinguished; Inputting the original sub-image group into the trained classification model to determine whether each original sub-image is a close-up image; The training steps of the classification model include: Collect the original puzzle picture; According to the jigsaw original image, each sub-image in each group of original sub-images is divided into different categories; According to each sub-image in each group of original sub-images, a trained classification model is obtained; The collecting of the original jigsaw puzzle image comprises: Collect original puzzle pictures of traffic violation evidence chains during the day and at night in different puzzle and close-up modes from various devices; According to the jigsaw original image, each sub-image in each group of original sub-images is divided into different categories, including: Splitting each of the jigsaw original images into a set of original sub-images; All original sub-images in each group of original sub-images are divided into different categories through manual annotation; All original sub-images in each group of original sub-images are divided into different categories through manual annotation, including: If the original sub-image is a close-up image, its corresponding category is marked as 0; If the original sub-image is a non-close-up image, its corresponding category is marked as 1; The step of obtaining a trained classification model according to each sub-image in each group of original sub-images includes: When the number of original sub-images in each group of original sub-images is less than 8, fill them with all-black fake sub-images; Each sub-image in each group of original sub-images is scaled to a fixed size, and then each scaled sub-image is horizontally spliced along its length to obtain a horizontal spliced image; the category corresponding to the completely black fake sub-image is marked as -1; Obtaining a trained classification model according to the category corresponding to each sub-image in the horizontal spliced image; The step of obtaining a trained classification model according to the category corresponding to each sub-image in the horizontal spliced image comprises: From left to right, set a label of the horizontal mosaic image according to the category corresponding to each sub-image in the horizontal mosaic image; Input the horizontal splicing image and its corresponding label into a deep convolutional neural network to extract features and obtain a 1*8 feature map; Use the obtained 1*8 feature map to train the pixel-level classification model to obtain a trained classification model; The step of determining whether each original sub-image is a close-up image includes: If the category corresponding to the original sub-image is 0, the original sub-image is judged as a close-up image; If the category corresponding to the original sub-image is 1, the original sub-image is judged as a non-close-up image.
2. The method according to claim 1, characterized in that The label corresponding to the all-black fake sub-image is -1; when training the classification model, the -1 label is ignored.
3. A close-up image recognition system based on deep learning, implementing the method of claim 1 or 2, characterized in that: include: A splitting module, used for splitting the puzzle to be distinguished into original sub-picture groups to be distinguished; The judgment module is used to input the original sub-image group into the trained classification model to judge whether each original sub-image is a close-up image.