Artificial intelligence-based target detection method and device

Abstract

The application discloses an object detection method and device based on artificial intelligence, and belongs to the field of graphic image processing.The specific steps of the detection method are as follows: (1) collecting and filtering target pictures; (2) performing feature extraction on the target pictures; (3) performing target recognition according to the feature information; (4) detecting and optimizing the recognition result; and (5) feeding back and storing the optimization scheme.The application can filter the redundant target pictures, greatly improves the target detection efficiency, meanwhile, the convolutional neural network is adopted to segment and extract the features of the target pictures, the information extraction integrity of the target pictures is improved, the error target detection result can be corrected automatically, the staff does not need to manually find, the staff workload is reduced, the work efficiency is improved, and the staff uses conveniently.

Description

Technical Field

[0001] This invention relates to the field of graphics and image processing, and more particularly to a target detection method and apparatus based on artificial intelligence. Background Technology

[0002] Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies, and application systems to simulate, extend, and expand human intelligence. It is a comprehensive technology within computer science that studies the design principles and implementation methods of various intelligent machines to enable them to possess perception, reasoning, and decision-making functions. It attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. With technological advancements, AI technology will be applied in more fields and play an increasingly important role.

[0003] A search revealed that Chinese patent CN112052837A discloses an artificial intelligence-based target detection method and device. While this invention improves the accuracy of target detection, it cannot filter out redundant images, reducing target detection efficiency, and it also cannot completely extract information from the target image. Furthermore, existing artificial intelligence-based target detection methods and devices require manual searching for erroneous detection results, reducing work efficiency and making them inconvenient for staff to use. Therefore, we propose an artificial intelligence-based target detection method and device. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based target detection method and apparatus.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An AI-based object detection method, the specific steps of which are as follows:

[0007] (1) Collect and filter target images: Collect one or more sets of target images and build a data filter. At the same time, import each set of target images into the data filter for data cleaning.

[0008] (2) Feature extraction of target images: Construct a convolutional neural network and import the filtered target images into the convolutional neural network for target feature extraction. At the same time, integrate and classify the extracted feature data into a feature dataset.

[0009] (3) Target recognition based on feature information: Artificial intelligence receives feature datasets, performs recognition analysis based on each group of feature data in the feature datasets, and records the recognition results of each group;

[0010] (4) Detect and optimize the recognition results: Collect the recognition results of each group, analyze their recognition accuracy, record the incorrect recognition results, and then conduct a second analysis on the incorrect recognition results, while optimizing and adjusting the original recognition scheme.

[0011] (5) Feedback and storage of optimization solutions: Feedback of recognition results to staff, and upload the optimized recognition solutions to the cloud database for storage and feedback.

[0012] As a further aspect of the present invention, the specific steps of data cleaning in step (1) are as follows:

[0013] Step 1: The computer acquires target images for each group through an external camera, and then the data filter classifies the target images according to the data source corresponding to each group of target images;

[0014] Step 2: Export the target images from different data sources into text files of a specified format or other file types, and then extract the target image files.

[0015] Step 3: The extracted target image files are processed by handling missing data, duplicate data, abnormal data, and inconsistent data to filter out unidentifiable or redundant target images in each group of collected target images.

[0016] As a further aspect of the present invention, the specific steps of feature extraction in step (2) are as follows:

[0017] Step 1: The convolutional neural network divides the obtained feature map into multiple local features according to different averaging criteria based on the topology corresponding to the target, and groups the local features obtained under the same criteria into the same group.

[0018] Step 2: Construct an adjacency matrix to enhance each group of local regions, and enable each local feature to learn information from other local features. At the same time, perform convolution operations on the upper and lower branches and then concatenate them.

[0019] The third step is to use the idea of ​​residual networks to fuse the features of each group of local regions to obtain a local feature enhanced by shallow relationships. At the same time, the local features of each group of the same target image are spliced ​​together to generate the corresponding global feature.

[0020] As a further aspect of the present invention, the specific steps of the identification and analysis in step (3) are as follows:

[0021] S1.1: Artificial intelligence extracts global features from the feature dataset in chronological order of collection time, while simultaneously establishing a communication connection with the Internet;

[0022] S1.2: Based on the global features of the target image, query relevant information from the data repository or the Internet, determine the object information corresponding to the target image, and construct an analysis record table to record the target image information and analysis results.

[0023] As a further aspect of the present invention, the accuracy analysis in step (4) is specifically performed as follows:

[0024] S2.1: Construct an analytical neural model and import the analytical record table into the analytical neural model. At the same time, the analytical neural model extracts information from each group in the analytical record table to construct a simulated dataset. Select a simulated dataset as validation data and use the validation data repeatedly to verify the accuracy of the analytical neural model.

[0025] S2.2: For each set of simulated data, select any subset as the test set, and then select the remaining subset as the training set. Make a prediction for each set of data and output the data with the best prediction result as the optimal parameter.

[0026] S2.3: Standardize the training dataset according to the optimal parameters, and finally send the training samples to the analysis neural model. At the same time, the trained analysis neural model is iteratively trained.

[0027] S2.4: Based on the training results of the neural model, plot the corresponding curves. At the same time, extract the target images with errors in target detection and the analysis results based on the curves, and integrate and classify the extracted data to generate an error correction dataset.

[0028] S2.5: Perform secondary detection analysis on the error correction dataset, and feed back the data that cannot be detected to the staff for manual analysis. At the same time, count the total number of detection targets, the number of detection errors, and the number of correct detections, and draw pie charts of the percentage of detection errors and the percentage of correct detections.

[0029] As a further aspect of the present invention, the specific steps for optimization and adjustment in step (4) are as follows:

[0030] P1: The optimization and update module collects and analyzes the curves generated by the neural model, extracts the original recognition scheme of artificial intelligence, and connects to the external server port for data.

[0031] P2: Simultaneously, normalize the collected data sets, draw relevant discrete models based on the processing results, and import the target data sets after error correction into the optimization and update module.

[0032] P3: The optimization and update module extracts one or more recognition schemes from the external server, tests their detection accuracy, and integrates the recognition scheme with the highest accuracy with the original recognition scheme. At the same time, it updates the target detection parameters in the original scheme based on the best parameters output by the analysis neural model, and replaces the original recognition scheme with the latest recognition scheme generated by integration.

[0033] The AI-based target detection device includes a data acquisition module, a feature extraction module, a target recognition module, a result analysis module, an optimization and update module, a cloud database, and a feedback display module.

[0034] The data acquisition module is used to communicate with an external camera and acquire images related to the target object.

[0035] The feature extraction module is used to receive the collected target images from each group, and simultaneously construct a convolutional neural network to extract features from each group of target images, while also constructing a relevant feature dataset.

[0036] The target recognition module is used to analyze each group of feature data in the feature dataset, and connects to the external Internet for communication, while collecting and recording the target information.

[0037] The result analysis module is used to collect target recognition results, analyze the accuracy of each group of recognition results, correct erroneous recognition results, and draw corresponding pie charts based on the analysis structure.

[0038] The optimization and update module is used to receive the original recognition scheme, recognition results and analysis results, and optimize and update the original recognition scheme.

[0039] The cloud database is used to receive updated identification schemes and store them in an orderly manner;

[0040] The feedback display module is used to receive target recognition results, error correction records, and optimized recognition schemes, and then feed them back to the staff.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. Compared with previous detection methods, this AI-based target detection method classifies target images into groups according to their corresponding data sources using a data filter. The target images from different data sources are then exported as text files or other file types in a specified format. The resulting target image files are then extracted. Further processing of the extracted target image files, including handling missing, duplicate, and abnormal data, as well as inconsistent data, filters out unidentifiable and redundant target images. Then, a convolutional neural network horizontally segments the feature map into multiple local features based on the target's corresponding topology and different averaging criteria. Local features obtained under the same criteria are grouped together. An adjacency matrix is ​​then constructed to enhance each group of local regions. This process effectively filters out redundant target images, significantly improving target detection efficiency. Simultaneously, the convolutional neural network segments and extracts target image features, enhancing the completeness of information extraction from the target images.

[0043] 2. This AI-based object detection method constructs an analytical neural model and imports an analytical record table into it. The model extracts information from each set of records to build a simulated dataset. One set of simulated data is selected as validation data and used repeatedly to verify the model's accuracy. A prediction is made for each set of data, and the best prediction is output as the optimal parameter. The training dataset is standardized based on the optimal parameter. Finally, training samples are fed back into the analytical neural model, which undergoes iterative training. The training results are then used to plot corresponding curves. Based on these curves, target images and analysis results with detection errors are extracted. The extracted data are integrated and categorized to generate an error correction dataset. This dataset undergoes secondary detection analysis, and data that cannot be detected is returned to staff for manual analysis. The method automatically corrects erroneous object detection results without requiring manual searching, reducing workload, improving efficiency, and simplifying operation. Attached Figure Description

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0045] Figure 1 This is a flowchart of the artificial intelligence-based target detection method proposed in this invention;

[0046] Figure 2 This is a system block diagram of the target detection device based on artificial intelligence proposed in this invention. Detailed Implementation

[0047] Example 1

[0048] Reference Figure 1 An artificial intelligence-based target detection method is described, and its specific steps are as follows:

[0049] Collect and filter target images: Collect one or more sets of target images, build a data filter, and import each set of target images into the data filter for data cleaning.

[0050] Specifically, the computer acquires target images from external cameras. A data filter then categorizes the target images according to their corresponding data sources and exports them as text files or other file types in a specified format. The resulting target image files are then extracted, and the extracted files are processed for missing, duplicate, and abnormal data, as well as inconsistent data, to filter out unidentifiable or redundant target images. This process significantly improves target detection efficiency by filtering out redundant images. Simultaneously, a convolutional neural network is used to segment and extract target image features, enhancing the completeness of information extraction from the target images.

[0051] Feature extraction of target images: Convolutional neural networks are constructed, and the filtered target images are imported into the convolutional neural networks for target feature extraction. At the same time, the extracted feature data are integrated and classified into a feature dataset.

[0052] Specifically, the convolutional neural network horizontally divides the obtained feature map into multiple local features according to different averaging criteria based on the topology corresponding to the target. Local features obtained under the same criteria are grouped into the same group. Then, an adjacency matrix is ​​constructed to enhance each group of local regions, and each local feature learns information from other local features. At the same time, convolution operations are performed on the upper and lower branches and concatenated. The idea of ​​residual network is used to fuse the features of each group of local regions to obtain a local feature enhanced by shallow relationship. At the same time, the local features of each group of the same target image are concatenated to generate the corresponding global feature.

[0053] Target recognition based on feature information: Artificial intelligence receives a feature dataset, performs recognition and analysis based on each group of feature data in the dataset, and records the recognition results for each group.

[0054] Specifically, artificial intelligence extracts global features from the feature dataset in chronological order of collection time, while communicating with the Internet. Based on the global features of the target image, it queries relevant information from the data repository or the Internet, determines the object information corresponding to the target image, and constructs an analysis record table to record the target image information and analysis results.

[0055] The recognition results are then tested and optimized: the recognition results of each group are collected, their recognition accuracy is analyzed, and incorrect recognition results are recorded. Then, the incorrect recognition results are analyzed again, and the original recognition scheme is optimized and adjusted.

[0056] Specifically, the computer constructs an analytical neural model and imports the analytical record table into it. Simultaneously, the analytical neural model extracts information from each group in the analytical record table to construct a simulated dataset. One simulated dataset is selected as validation data, and this validation data is used repeatedly to verify the accuracy of the analytical neural model. For each set of simulated data, an arbitrary subset is selected as the test set, and the remaining subset is used as the training set. A prediction is made for each set of data, and the data with the best prediction result is output as the optimal parameters. The training dataset is standardized based on the optimal parameters. Finally, the training samples are fed into the analytical neural model, and the trained analytical neural model... The model undergoes iterative training, and then the corresponding curves are plotted based on the training results of the neural model. Simultaneously, based on the curves, target images with detection errors and analysis results are extracted, and the extracted data are integrated and categorized to generate an error correction dataset. The error correction dataset undergoes secondary detection analysis, and data that cannot be detected is fed back to staff for manual analysis. The system also counts the total number of detected targets, the number of detection errors, and the number of correct detections, and plots pie charts showing the percentage of detection errors and correct detections. It can automatically correct erroneous target detection results without requiring manual searching by staff, reducing workload, improving efficiency, and making it convenient for staff to use.

[0057] Specifically, the optimization and update module collects and analyzes the curves generated by the neural model, extracts the original AI recognition schemes, connects to the external server port, normalizes the collected data, draws relevant discrete models based on the processing results, and imports the target data after error correction into the optimization and update module. Then, the optimization and update module extracts one or more recognition schemes from the external server, tests their detection accuracy, and integrates the recognition scheme with the highest accuracy with the original recognition scheme. At the same time, it updates the target detection parameters in the original scheme based on the optimal parameters output by the neural model, replacing the original recognition scheme with the latest integrated recognition scheme.

[0058] Feedback and storage of optimization solutions: The recognition results are fed back to the staff, and the optimized recognition solutions are uploaded to the cloud database for storage and feedback.

[0059] Example 2

[0060] Reference Figure 2 The target detection device based on artificial intelligence includes a data acquisition module, a feature extraction module, a target recognition module, a result analysis module, an optimization and update module, a cloud database, and a feedback display module.

[0061] The data acquisition module is used to communicate with an external camera and acquire images related to the target object.

[0062] The feature extraction module is used to receive the collected target images from each group, and at the same time construct a convolutional neural network to extract features from each group of target images, and construct a relevant feature dataset.

[0063] The target recognition module is used to analyze each group of feature data in the feature dataset, and communicates with the external Internet to collect and record the target information.

[0064] The results analysis module is used to collect target recognition results, analyze the accuracy of each group of recognition results, correct erroneous recognition results, and draw corresponding pie charts based on the analysis structure.

[0065] The optimization and update module is used to receive the original recognition scheme, recognition results and analysis results, and at the same time optimize and update the original recognition scheme.

[0066] The cloud database is used to receive updated recognition schemes and store them in an orderly manner;

[0067] The feedback display module is used to receive target recognition results, error correction records, and optimized recognition schemes, and then feed them back to the staff.

Claims

1. An artificial intelligence-based target detection method, characterized in that, The specific steps of this detection method are as follows: (1) Collect and filter target images: Collect one or more sets of target images and build a data filter. At the same time, import each set of target images into the data filter for data cleaning. (2) Feature extraction of target images: Construct a convolutional neural network and import the filtered target images into the convolutional neural network for target feature extraction. At the same time, integrate and classify the extracted feature data into a feature dataset. (3) Target recognition based on feature information: Artificial intelligence receives feature datasets, performs recognition analysis based on each group of feature data in the feature datasets, and records the recognition results of each group; (4) Detect and optimize the recognition results: Collect the recognition results of each group, analyze their recognition accuracy, record the incorrect recognition results, and then conduct a second analysis of the incorrect recognition results, while optimizing and adjusting the original recognition scheme. (5) Feedback and storage of optimization solutions: Feedback of recognition results to staff, and upload the optimized recognition solution to the cloud database for storage and feedback; The specific steps of data cleaning described in step (1) are as follows: Step 1: The computer acquires target images for each group through an external camera, and then the data filter classifies the target images according to the data source corresponding to each group of target images; Step 2: Export the target images from different data sources into text files of a specified format or other file types, and then extract the target image files. Step 3: The extracted target image files are processed by handling missing data, duplicate data, abnormal data, and inconsistent data to filter out unidentifiable and redundant target images in each group of collected target images. The specific steps of feature extraction described in step (2) are as follows: Step 1: The convolutional neural network divides the obtained feature map into multiple local features according to different averaging criteria based on the topology corresponding to the target, and groups the local features obtained under the same criteria into the same group. Step 2: Construct an adjacency matrix to enhance each group of local regions, and enable each local feature to learn information from other local features. At the same time, perform convolution operations on the upper and lower branches and then concatenate them. Step 3: Using the idea of ​​residual networks, the features of each group of local regions are fused to obtain a local feature enhanced by shallow relationships. At the same time, the local features of each group of the same target image are spliced ​​together to generate the corresponding global feature. The specific steps of the identification and analysis described in step (3) are as follows: S1.1: Artificial intelligence extracts global features from the feature dataset in chronological order of collection time, while simultaneously establishing a communication connection with the Internet; S1.2: Based on the global features of the target image, query relevant information from the data repository or the Internet, determine the object information corresponding to the target image, and construct an analysis record table to record the target image information and analysis results; The specific steps of the accuracy analysis described in step (4) are as follows: S2.1: Construct an analytical neural model and import the analytical record table into the analytical neural model. At the same time, the analytical neural model extracts information from each group in the analytical record table to construct a simulated dataset. Select a simulated dataset as validation data and use the validation data repeatedly to verify the accuracy of the analytical neural model. S2.2: For each set of simulated data, select any subset as the test set, and then select the remaining subset as the training set. Make a prediction for each set of data and output the data with the best prediction result as the optimal parameter. S2.3: Standardize the training dataset according to the optimal parameters, and finally send the training samples to the analysis neural model. At the same time, the trained analysis neural model is iteratively trained. S2.4: Based on the training results of the neural model, plot the corresponding curves. At the same time, extract the target images with errors in target detection and the analysis results based on the curves, and integrate and classify the extracted data to generate an error correction dataset. S2.5: Perform secondary detection analysis on the error correction dataset, and feed back the data that cannot be detected to the staff for manual analysis. At the same time, count the total number of detection targets, the number of detection errors and the number of correct detections, and draw a pie chart of the percentage of detection errors and the percentage of correct detections. The specific steps for optimization and adjustment described in step (4) are as follows: P1: The optimization and update module collects and analyzes the curves generated by the neural model, extracts the original recognition scheme of artificial intelligence, and connects to the external server port for data. P2: Simultaneously, normalize the collected data sets, draw relevant discrete models based on the processing results, and import the target data sets after error correction into the optimization and update module. P3: The optimization and update module extracts one or more recognition schemes from the external server, tests their detection accuracy, and integrates the recognition scheme with the highest accuracy with the original recognition scheme. At the same time, it updates the target detection parameters in the original scheme based on the best parameters output by the analysis neural model, and replaces the original recognition scheme with the latest recognition scheme generated by integration.

2. A target detection device based on artificial intelligence, characterized in that, It includes a data acquisition module, a feature extraction module, a target recognition module, a result analysis module, an optimization and update module, a cloud database, and a feedback display module; The data acquisition module is used to communicate with an external camera and acquire images related to the target object. The feature extraction module is used to receive the collected target images from each group, and simultaneously construct a convolutional neural network to extract features from each group of target images, while also constructing a relevant feature dataset. The target recognition module is used to analyze each group of feature data in the feature dataset, and connects to the external Internet for communication, while collecting and recording the target information. The result analysis module is used to collect target recognition results, analyze the accuracy of each group of recognition results, correct erroneous recognition results, and draw corresponding pie charts based on the analysis structure. The optimization and update module is used to receive the original recognition scheme, recognition results and analysis results, and optimize and update the original recognition scheme. The cloud database is used to receive updated identification schemes and store them in an orderly manner; The feedback display module is used to receive target recognition results, error correction records, and optimized recognition schemes, and then feed them back to the staff. The specific steps of the feature extraction module in constructing a convolutional neural network to extract features from each group of target images are as follows: Step 1: The convolutional neural network divides the obtained feature map into multiple local features according to different averaging criteria based on the topology corresponding to the target, and groups the local features obtained under the same criteria into the same group. Step 2: Construct an adjacency matrix to enhance each group of local regions, and enable each local feature to learn information from other local features. At the same time, perform convolution operations on the upper and lower branches and then concatenate them. Step 3: Using the idea of ​​residual networks, the features of each group of local regions are fused to obtain a local feature enhanced by shallow relationships. At the same time, the local features of each group of the same target image are spliced ​​together to generate the corresponding global feature. The specific steps for the result analysis module to analyze the accuracy of each group of identification results are as follows: S2.1: Construct an analytical neural model and import the analytical record table into the analytical neural model. At the same time, the analytical neural model extracts information from each group in the analytical record table to construct a simulated dataset. Select a simulated dataset as validation data and use the validation data repeatedly to verify the accuracy of the analytical neural model. S2.2: For each set of simulated data, select any subset as the test set, and then select the remaining subset as the training set. Make a prediction for each set of data and output the data with the best prediction result as the optimal parameter. S2.3: Standardize the training dataset according to the optimal parameters, and finally send the training samples to the analysis neural model. At the same time, the trained analysis neural model is iteratively trained. S2.4: Based on the training results of the neural model, plot the corresponding curves. At the same time, extract the target images with errors in target detection and the analysis results based on the curves, and integrate and classify the extracted data to generate an error correction dataset. S2.5: Perform secondary detection analysis on the error correction dataset, and feed back the data that cannot be detected to the staff for manual analysis. At the same time, count the total number of detection targets, the number of detection errors and the number of correct detections, and draw a pie chart of the percentage of detection errors and the percentage of correct detections. The specific steps for the optimization and update module to optimize and update the original recognition scheme are as follows: P1: The optimization and update module collects and analyzes the curves generated by the neural model, extracts the original recognition scheme of artificial intelligence, and connects to the external server port for data. P2: Simultaneously, normalize the collected data sets, draw relevant discrete models based on the processing results, and import the target data sets after error correction into the optimization and update module. P3: The optimization and update module extracts one or more recognition schemes from the external server, tests their detection accuracy, and integrates the recognition scheme with the highest accuracy with the original recognition scheme. At the same time, it updates the target detection parameters in the original scheme based on the best parameters output by the analysis neural model, and replaces the original recognition scheme with the latest recognition scheme generated by integration.

Images