Target identification method based on template matching
By using generative adversarial networks in target recognition to generate elastic templates and combining multi-level matching technology, the problems of complex computing, limited generalization and poor robustness in the existing technology are solved, and the target recognition effect with high precision and robustness is achieved.
Patent Information
- Application Number
- CN202510025851.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-08
- Publication Date
- 2025-04-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art has problems in target recognition with high computational complexity, limited generalization ability and poor robustness, especially in the case of lighting changes, occlusion and complex background.
A target recognition method based on template matching is adopted to generate elastic templates containing different morphology and partial occlusion conditions of the target object by generating an adversarial network (GAN), and multi-level matching and template updates are performed by combining color histograms, fast feature matching, multi-scale matching and depth feature matching.
It improves the accuracy and robustness of target recognition, reduces the computing burden and storage overhead, enhances the system's adaptability and self-learning ability, and ensures high-precision recognition in different complex environments.
Smart Images
Figure CN119850984A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer vision, and in particular to a target recognition method based on template matching. Background Art
[0002] The target recognition methods in the prior art usually rely on traditional computer vision technology and deep learning algorithms. The specific implementation methods include methods based on feature matching, which involve feature extraction, such as using SIFT, SURF, ORB, etc. to extract key points and descriptors. Feature matching determines the target position by matching the key points of the input image and the template image. Algorithms such as FLANN and RANSAC are often used, and post-processing is performed to improve accuracy; methods based on deep learning need to collect image data from multiple angles and lighting conditions to construct a training set, use CNN or other deep learning models for training, and realize target detection through model forward propagation. Common models include YOLO, Faster R-CNN, SSD, etc., which also require post-processing to improve the accuracy of the results; there are also methods based on template matching, which generate templates by taking images of the target object at different angles, use template matching algorithms to search for similar areas in the input image, and use multi-scale matching, overlapping area merging and other technologies to optimize the matching results.
[0003] The existing technologies have the following disadvantages when achieving target recognition: for methods based on feature matching, the recognition effect is highly dependent on the quality of the extracted features and the performance of the matching algorithm. In the case of illumination changes, occlusion, complex background, etc., feature extraction and matching may fail, and the computational complexity is high. Feature matching usually requires a large amount of computing resources, especially in large-scale images or real-time applications, where the computing time is too long; for methods based on deep learning, large-scale annotated data sets are required for training, and the data collection and annotation process is time-consuming and costly. Although the model performs well on the training set, it has limited generalization ability in unseen environments and occlusions, and is prone to false detection or missed detection. In addition, deep learning models usually require high-performance GPUs for real-time reasoning, which places high demands on computing resources; for methods based on template matching, traditional template generation is fixed and cannot adapt to different forms and occlusion conditions of target objects, resulting in low recognition accuracy. In order to improve the recognition accuracy, a large number of templates need to be generated, which increases storage and computing overhead. In the case of illumination changes, complex backgrounds and partial occlusions, the robustness is poor, and false matching is prone to occur. Summary of the invention
[0004] In view of the deficiencies of the prior art, the present invention provides a target recognition method based on template matching, which solves the problems of complex calculation and limited generalization in the prior art.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a target recognition method based on template matching, specifically comprising the following steps: S1. Data collection Collect image data of target objects from different angles and lighting conditions; S2. Data processing Preprocess the collected image data, including grayscale conversion, size unification and normalization; S3. Template creation Generate elastic templates that contain different shapes and partial occlusions of the target object using a generative adversarial network; S4. Preliminary screening The input image is preliminarily screened by using color histogram and fast feature matching methods to exclude areas with obvious mismatches; S5. Multi-level matching Perform multi-level matching, including multi-scale matching and deep feature matching, to ultimately determine the position and posture of the target object; S6. Template Update The template library is updated according to the matching results to improve the recognition accuracy and robustness of the system.
[0006] Preferably, in the step S2 data processing, the quality of the generated templates is evaluated by a discriminator, and high-quality templates are selected and stored in the template library.
[0007] Preferably, the preliminary screening in step S4 includes using SIFT or SURF feature matching method to further improve the accuracy of the preliminary screening.
[0008] Preferably, the multi-level matching in step S5 includes dynamically adjusting the matching strategy in combination with the context information of the input image to improve the accuracy and robustness of the matching.
[0009] Preferably, the data processing in step S2 includes performing random cropping, rotation and scaling enhancement techniques on the image to increase the diversity and robustness of the data, normalizing and removing noise from the depth image to improve the extraction accuracy of the depth feature.
[0010] Preferably, the multi-level matching in step S5 includes using a local feature matching method to improve the recognition accuracy in the local area of the target object.
[0011] Preferably, the template updating in step S6 includes determining the templates that need to be updated through manual annotation or automatic detection methods, and regularly evaluating the quality of the template library to eliminate low-quality templates to maintain the efficiency and accuracy of the template library.
[0012] The present invention provides a target recognition method based on template matching. It has the following beneficial effects: The present invention provides a target recognition method based on template matching. The present invention generates elastic templates containing different forms and partial occlusion conditions of target objects by using a generative adversarial network (GAN), thereby enhancing adaptability. The quality of the generated templates is evaluated by a discriminator, and the templates are selected and stored in a template library, thereby improving the representativeness of the templates. Preliminary screening is performed by a color histogram and a fast feature matching method, and areas that are obviously not matched are quickly excluded, thereby reducing the subsequent calculation burden. Multi-scale matching and deep feature matching, as well as context-aware matching, a multi-level strategy ensures high-precision recognition in different complex environments. The matching strategy is dynamically adjusted, and context information is combined to improve the robustness and accuracy of recognition. Error results of recognition are fed back to a generator for incremental learning, and only some model parameters are updated, thereby improving adaptability and self-learning ability. The quality of the template library is regularly evaluated, new templates are added or invalid templates are removed, and the efficiency and accuracy of the template library are maintained, thereby reducing storage and calculation overheads. By collecting user feedback, the system function and performance are further optimized, thereby improving user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0014] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.
[0015] like Figure 1 As shown, the embodiment of the present invention provides a target recognition method based on template matching, which specifically includes the following steps: S1. Data collection Collect image data of target objects from different angles and lighting conditions; S2. Data processing Preprocess the collected image data, including grayscale conversion, size unification and normalization; S3. Template creation Generate elastic templates that contain different shapes and partial occlusions of the target object using a generative adversarial network; S4. Preliminary screening The input image is preliminarily screened by using color histogram and fast feature matching methods to exclude areas with obvious mismatches; S5. Multi-level matching Perform multi-level matching, including multi-scale matching and deep feature matching, to ultimately determine the position and posture of the target object; S6. Template Update The template library is updated according to the matching results to improve the recognition accuracy and robustness of the system.
[0016] In the data processing of step S2, the quality of the generated template is evaluated by the discriminator, and high-quality templates are selected and stored in the template library. The preliminary screening of step S4 includes using SIFT or SURF feature matching methods to further improve the accuracy of the preliminary screening. The multi-level matching of step S5 includes dynamically adjusting the matching strategy in combination with the contextual information of the input image to improve the accuracy and robustness of the matching. The data processing of step S2 includes performing random cropping, rotation and scaling enhancement techniques on the image to increase the diversity and robustness of the data, normalizing and removing noise from the depth image, and improving the extraction accuracy of the depth feature. The multi-level matching of step S5 includes using a local feature matching method to improve the recognition accuracy in the local area of the target object. The template update of step S6 includes determining the template that needs to be updated through manual annotation or automatic detection methods, and regularly evaluating the quality of the template library, eliminating low-quality templates, and maintaining the efficiency and accuracy of the template library.
[0017] Specifically: Data collection and preprocessing Use a high-definition camera to collect images of the target object from different angles and lighting conditions. Use an infrared camera to collect infrared images of the target object under different temperature and environmental conditions. Use a depth camera to collect 3D depth information of the target object. Use a high-sensitivity microphone to collect audio signals around the target object under different environmental noise conditions.
[0018] Convert color images to grayscale to reduce data dimensions and improve computational efficiency. Resize all images to a uniform size, such as 256x256 pixels. Normalize image pixel values to between 0 and 1 for model training. Increase data diversity and robustness through random cropping, rotation, and scaling techniques.
[0019] Adjust the depth value to a uniform range, such as between 0 and 1. Use median filtering or Gaussian filtering to remove noise from the depth image and improve the template quality.
[0020] Use a bandpass filter to remove background noise and retain the audio features of the target object. Convert the audio signal into a spectrogram and extract the spectral features for template generation and subsequent matching.
[0021] Flexible template generation The generator generates flexible templates that contain different shapes and partial occlusions of the target object. The generator generates these templates by learning input data (preprocessed visual images, depth images, and audio features). The discriminator is used to evaluate whether the generated templates are realistic. The discriminator helps the generator generate higher quality templates by learning the difference between real templates and generated templates.
[0022] Build the generator and discriminator model structures. Load the training data, including images, depth images, and audio features of real target objects. The generator uses random noise as input to generate templates. The discriminator evaluates whether the generated template is realistic and feeds the result back to the generator. Based on the feedback from the discriminator, update the parameters of the generator and discriminator. Repeat the above steps until the generator can generate high-quality elastic templates.
[0023] The preprocessed visual image, depth image and audio features are used as input to the generator. The generator is used to generate elastic templates that contain different forms and partial occlusions of the target object. The quality of the generated templates is evaluated by the discriminator, and high-quality templates are selected and stored in the template library.
[0024] Template Matching Calculate the color histogram of the input image and the template, and preliminarily screen out the areas with high matching degree through histogram comparison. Use ORB and FAST fast feature matching methods to quickly exclude areas with obvious mismatches.
[0025] Perform template matching at different scales of the input image to find the rough target location. By scaling the image and template, targets of different sizes can be captured. The generator generates local feature templates for more refined matching in specific areas. Use a 3D convolutional neural network to extract deep features of the input image and template for more accurate matching. Combine the contextual information of the input image (such as surrounding objects, environmental conditions), dynamically adjust the matching strategy, and improve the accuracy and robustness of the matching.
[0026] Template Update and Maintenance Feedback the recognized error results to the generator to help it learn new template generation strategies. The generator performs incremental learning based on the feedback data and only updates some model parameters to adapt to new data and environmental changes.
[0027] The generated high-quality templates are stored in the template library for subsequent matching. The template library is updated regularly or as needed to add new templates or remove invalid templates. Templates that need to be updated can be determined by manual annotation or automatic detection methods.
[0028] System integration and testing The data acquisition and preprocessing module is used to collect and preprocess data to ensure the consistency and quality of input data. The elastic template generation module is set up to generate high-quality elastic templates and provide diversified matching templates. The template matching module is used to perform preliminary screening, multi-level matching, and adversarial matching to ensure the accuracy and robustness of recognition. The template update and maintenance module is used to manage the template library and incremental learning to ensure that the system can adapt to new data and environments.
[0029] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A target recognition method based on template matching, characterized in that: The specific steps include: S1. Data collection Collect image data of target objects from different angles and lighting conditions; S2. Data processing Preprocess the collected image data, including grayscale conversion, size unification and normalization; S3. Template creation Generate elastic templates that contain different shapes and partial occlusions of the target object using a generative adversarial network; S4. Preliminary screening The input image is preliminarily screened by using color histogram and fast feature matching methods to exclude areas with obvious mismatches; S5. Multi-level matching Perform multi-level matching, including multi-scale matching and deep feature matching, to ultimately determine the position and posture of the target object; S6. Template Update The template library is updated according to the matching results to improve the recognition accuracy and robustness of the system.
2. The target recognition method based on template matching according to claim 1, characterized in that: In the step S2 data processing, the quality of the generated template is evaluated by the discriminator, and the template is selected and stored in the template library.
3. The target recognition method based on template matching according to claim 1, characterized in that: The preliminary screening in step S4 includes using SIFT or SURF feature matching method to further improve the accuracy of the preliminary screening.
4. The target recognition method based on template matching according to claim 1, characterized in that: The multi-level matching in step S5 includes dynamically adjusting the matching strategy in combination with the context information of the input image.
5. The target recognition method based on template matching according to claim 1, characterized in that: The data processing in step S2 includes performing random cropping, rotation and scaling enhancement techniques on the image, and performing normalization and noise removal processing on the depth image.
6. The target recognition method based on template matching according to claim 1, characterized in that: The multi-level matching in step S5 includes using a local feature matching method to improve the recognition accuracy in the local area of the target object.
7. The target recognition method based on template matching according to claim 1, characterized in that: The template updating in step S6 includes determining the templates that need to be updated through manual annotation or automatic detection methods, and regularly evaluating the quality of the template library to eliminate templates with low quality.