Image recognition method
By analyzing the target characteristics of the preset relationship in the image recognition method, the computing power consumption and sample acquisition problems of massive sample training in the existing technologies are solved, and efficient image recognition and accuracy are improved.
Patent Information
- Application Number
- CN202411362538.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2025-01-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
When existing image recognition methods deal with a single scene with specific rules, they require massive sample training, resulting in high computing power consumption and time-consuming and labor-intensive acquisition of samples.
A method of image recognition is proposed. By obtaining image samples and training models, the first target characteristics and second target characteristics with preset relationships are used to save computing power and improve the accuracy of identification.
This method can improve the accuracy of image recognition under the condition that computing power is saved, and further improves accuracy through subsequent sample learning.
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition, and in particular to an image recognition method. Background Art
[0002] Existing image recognition methods are mainly divided into analytical methods and model methods. Although the model method can make the analysis results more and more accurate after a large number of sample training and has strong learning ability, for some single scenes with specific rules, the method of improving the accuracy of analysis results through massive sample training is relatively computationally intensive, and the acquisition of massive samples is also time-consuming and labor-intensive, which urgently needs to be improved. Summary of the invention
[0003] In order to solve the technical problems existing in the background technology, the present invention proposes an image recognition method, comprising:
[0004] Obtaining image samples, and using the image samples to train the model, wherein a positive sample in the image samples includes a first target feature and a second target feature having a preset association relationship;
[0005] The trained model is used to determine whether a target image having a first target feature contains a second target feature having a preset association relationship with the first target feature.
[0006] The first target feature and the second target feature having a preset association relationship specifically include: the first target feature has a first posture, the second target feature has a second posture, and there is a preset positional relationship between the first target feature and the second target feature.
[0007] Determining whether the target image contains a plurality of first target features and second target features having a preset association relationship, further comprising:
[0008] A first image including a first target feature and the first target feature being in a first posture is acquired from multiple frames of images, and the target image can be obtained by preprocessing the first image.
[0009] Preprocessing the first image specifically includes: cutting out a frame image containing the first target feature from the first image according to a preset rule.
[0010] Acquiring a first image including a first target feature and the first target feature being in a first posture from multiple frames of images specifically includes:
[0011] Define the first type of key points of the first target individual; establish a coordinate change curve of the first type of key points during movement; obtain all images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same movement trend as the first image.
[0012] The first target feature belongs to the first target individual, and the second target feature belongs to the target equipment. Preferably, the first target individual is a rope skipping individual, and the target equipment is a rope skipping.
[0013] All images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same motion trend are all first images. Specifically, all images in which the first type of key points are at the same vertical coordinate and have the same motion trend are all first images. Preferably, the first type of key points are specifically the center points of both shoulders, the motion trend is specifically an upward trend or a downward trend, and the first target feature has at least the wrist bone key point of the first target individual.
[0014] The first target feature belongs to the first target individual, and the second target feature belongs to the second target individual. Preferably, the first target individual is a pull-up exercise individual.
[0015] All images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same motion trend are all first images. Specifically, all images in which the first type of key points are at the same vertical coordinate and have the same motion trend are all first images. Preferably, the motion trend is an upward trend or a downward trend. Preferably, the first type of key points are shoulder points. Preferably, the first target feature has at least one or more contact points between the first target individual and the second target individual.
[0016] In the present invention, the image recognition method proposed can save computing power and at the same time, the recognition accuracy can be continuously improved. DETAILED DESCRIPTION
[0017] In order to solve the problems existing in the background technology, the present invention proposes an image recognition method, including an image recognition method, comprising:
[0018] An image sample is obtained and the model is trained using the image sample, wherein a positive sample in the image sample includes a first target feature and a second target feature having a preset association relationship.
[0019] The negative sample in the image sample may include the first target feature or the second target feature, and the negative sample in the image sample may also include the first target feature and the second target feature, but the first target feature and the second target feature do not have a preset association relationship.
[0020] The trained model is used to determine whether the target image contains a first target feature and a second target feature having a preset association relationship.
[0021] Since only the first target feature and the second target feature and the preset association relationship between the first target feature and the second target feature are used to analyze the image, computing power can be saved and reaction speed can be improved. At the same time, in some special scenarios, this method is sufficient to meet the accuracy, and through continuous subsequent sample learning, the accuracy can be further improved.
[0022] Furthermore, the first target feature has a first posture, the second target feature has a second posture, and there is a preset positional relationship between the first target feature and the second target feature. In other words, the first target feature with the first posture and the second target feature with the second posture have a preset positional relationship, that is, the first target feature and the second target feature have a preset association relationship.
[0023] Furthermore, before determining whether the target image contains multiple first target features and second target features having a preset association relationship, the method also includes: obtaining a first image containing the first target feature and the first target feature in a first posture from multiple frames of images, and preprocessing the first image to obtain the target image.
[0024] Preprocessing the first image specifically includes cutting out a frame diagram containing the first target feature from the first image according to a preset rule. Cutting out a frame diagram containing only useful information and inputting it into the model can save computing power.
[0025] Acquiring a first image including a first target feature and the first target feature being in a first posture from multiple frames of images specifically includes:
[0026] Further explanation is given by the rope skipping scenario:
[0027] A first type of key points of the first target individual is defined.
[0028] Among them, the first target feature belongs to the first target individual, and the second target feature belongs to the target equipment. The first target individual is specifically a skipping rope individual, and the target equipment is specifically a skipping rope.
[0029] A coordinate change curve of the first type of key points during the movement is established. Further, the coordinate change curve is a height change curve.
[0030] Rope skipping is a cyclic reciprocating motion. During multiple rope skippings, the first type of key point goes back and forth up and down, and the height change curve is a wave-like curve.
[0031] Specifically, the first type of key points are shoulder points, and the first target feature has at least the wrist key points of the first target individual. The wrist key points are easier to identify based on the skeletal key point recognition model. In addition, the wrist key points are relatively close to the held part of the skipping rope, and during skipping, the local speed of the skipping rope near the wrist is slower, so the features are easier to collect clearly, and the positive sample can only include this part of the skipping rope, so that a smaller frame diagram that meets the requirements can be cut out, thereby reducing the computing power occupied and improving the accuracy of the judgment results.
[0032] All images in which the first type of key points are at the same height and have the same motion trend are obtained as first images.
[0033] The height change curve has multiple repeated quasi-sine curves. In the same sinusoidal curve interval, the same height has the opportunity to correspond to a first type point and a second type point at the same time. Assuming that at the first type point moment, the first target individual is in the upward jumping stage, then the first type key point is in an upward trend; at the second type point moment, the first target individual is in the falling stage, then the second type key point is in a downward trend, then all images collected at the first type point moment or all images collected at the second type point moment can be selected as the first image. It should be pointed out that the image samples must also be collected according to the same rules.
[0034] In the normal rope skipping process, the first target individual, that is, the rope skipping individual, is the carrier of the first target feature, and the sports equipment, that is, the rope skipping, is the carrier of the second target feature. Since the posture of the individual athlete is correlated with the posture of the first target individual and their positions are also correlated, in the rope skipping scenario, this method can be used to identify whether there is a violation of ropeless jumping.
[0035] Furthermore, the following is further explained by taking how to identify the violation that the second target individual assists the first target individual in performing a pull-up exercise by hugging the first target individual and applying an upward thrust as an example:
[0036] The first target feature belongs to the first target individual, and the second target feature belongs to the second target individual. Preferably, the first target individual is a pull-up exercise individual.
[0037] All images where the first type of key points are located at the same horizontal coordinate or the same vertical coordinate and have the same motion trend are obtained as first images:
[0038] All images in which the first type of key points are at the same vertical coordinate and have the same movement trend are the first images. The first type of key points may be the jaw key points, etc. Based on the previous description, it will not be repeated here.
[0039] Preferably, the movement trend is specifically an upward trend or a downward trend. Preferably, the first target feature has at least one or more contact points between the first target individual and the second target individual.
[0040] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. An image recognition method, characterized in that: include: Obtaining image samples, and using the image samples to train the model, wherein a positive sample in the image samples includes a first target feature and a second target feature having a preset association relationship; The trained model is used to determine whether a target image having a first target feature contains a second target feature having a preset association relationship with the first target feature.
2. The image recognition method according to claim 1, characterized in that: The first target feature and the second target feature having a preset association relationship specifically include: the first target feature has a first posture, the second target feature has a second posture, and there is a preset positional relationship between the first target feature and the second target feature.
3. The image recognition method according to claim 2, characterized in that: Determining whether the target image contains a plurality of first target features and second target features having a preset association relationship, further comprising: A first image including a first target feature and the first target feature being in a first posture is acquired from multiple frames of images, and the target image can be obtained by preprocessing the first image.
4. The image recognition method according to claim 3, characterized in that: Preprocessing the first image specifically includes: cutting out a frame image containing the first target feature from the first image according to a preset rule.
5. The image recognition method according to claim 3, characterized in that: Acquiring a first image including a first target feature and the first target feature being in a first posture from multiple frames of images specifically includes: Define the first type of key points of the first target individual; establish a coordinate change curve of the first type of key points during movement; obtain all images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same movement trend as the first image.
6. The image recognition method according to claim 2, characterized in that: The first target feature belongs to the first target individual, and the second target feature belongs to the target equipment. Preferably, the first target individual is a rope skipping individual, and the target equipment is a rope skipping.
7. The image recognition method according to claim 5 or 6, characterized in that: All images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same motion trend are all first images. Specifically, all images in which the first type of key points are at the same vertical coordinate and have the same motion trend are all first images. Preferably, the first type of key points are specifically the center points of both shoulders, the motion trend is specifically an upward trend or a downward trend, and the first target feature has at least the wrist bone key point of the first target individual.
8. The image recognition method according to claim 2, characterized in that: The first target feature belongs to the first target individual, and the second target feature belongs to the second target individual. Preferably, the first target individual is a pull-up exercise individual.
9. The image recognition method according to claim 5 or 8, characterized in that: All images in which the first type of key points are at the same horizontal coordinate or the same vertical coordinate and have the same motion trend are all first images. Specifically, all images in which the first type of key points are at the same vertical coordinate and have the same motion trend are all first images. Preferably, the motion trend is an upward trend or a downward trend. Preferably, the first type of key points are shoulder points. Preferably, the first target feature has at least one or more contact points between the first target individual and the second target individual.