Method for identifying important parts by adopting region perception

Abstract

The invention discloses an important part identification method adopting area perception, and the method comprises the steps: directly carrying out the detailed attention distribution of an area where a target is located through employing the obtained perception information on the basis of the preliminary completion of target identification; and then, the self-adaptive attention module dynamically adjusts the working gravity center of the self-adaptive attention module, carries out deep analysis on key parts in the target area, and effectively captures and distinguishes the characteristics of the parts, thereby realizing high-precision key part identification. The method not only reduces unnecessary calculation overhead, but also ensures the accuracy and robustness of key part identification. By optimizing the calculation process and intelligent resource allocation, an efficient target and key part identification scheme is provided for an embedded system and other resource-limited scenes, the intelligent level in the fields of intelligent monitoring, automatic driving, medical image analysis and the like is expected to be improved, and intelligent development and application of the industry are promoted.

Description

Technical Field

[0001] The invention belongs to the technical field of pattern recognition, and in particular relates to a method for identifying key parts using regional perception. Background Art

[0002] With the in-depth evolution of digital transformation, the field of target detection and recognition is undergoing a revolutionary leap, from simple target recognition to precise positioning and deep structural analysis, marking a significant expansion of the scope of technology application. This process not only broadens the application prospects of technology in security monitoring, industrial automation, medical image analysis and other fields, but also heralds a new era of cross-industry collaborative innovation.

[0003] At present, a series of advanced algorithms have demonstrated excellent performance in target detection and classification in complex visual environments, and have built a solid foundation for the data processing process. The current target part recognition technology integrates instance segmentation and semantic segmentation technology on the basis of recognition to achieve accurate division of target entities with pixel-level accuracy, while giving each pixel a clear category label. The structural characteristics, state evaluation and functional analysis of the target are realized. With the help of deep learning technology, target part recognition technology can accurately mark key points or specific areas, so as to achieve good results in the keen capture of tiny defects in industrial detection and the identification and analysis of threat sources in the field of intelligent security.

[0004] However, the challenges encountered in actual deployment cannot be ignored, especially the high computing resource requirements, the limitations of real-time processing capabilities, and the interference of complex external factors, which have become key factors restricting the popularization and application of technology. In a resource-constrained environment, the reliance on high-performance AI computing platforms not only significantly raises the threshold of hardware costs, but also directly hinders the widespread application of technology. In addition, the interference of complex factors, such as changes in lighting conditions, has a serious impact on the segmentation effect, further highlighting the fragility and limitations of the technology in real-world applications.

[0005] Disadvantages of existing technology:

[0006] (1) High computing resource consumption: Current target detection and recognition technologies face high computing resource requirements when deployed in practice, especially in resource-constrained environments such as embedded systems. This not only increases hardware costs but also limits the widespread application of the technology.

[0007] (2) Limited real-time processing capabilities: The technology is not capable of real-time response in processing high-dimensional data and dynamic environments, making it difficult to meet the needs of scenarios such as real-time monitoring and autonomous driving.

[0008] (3) Sensitive to complex external factors: When dealing with complex factors such as illumination changes, occlusion, and background clutter, the segmentation effect and recognition accuracy of existing technologies are significantly reduced, exposing their fragility and limitations in real-world applications.

[0009] (4) Non-optimal resource allocation: Traditional methods usually perform calculations on the entire image and fail to intelligently focus on the target area, resulting in resource waste and inefficient processing.

[0010] (5) Difficulty in identifying key parts: In complex scenarios, it is challenging to accurately identify and distinguish the key parts of a target, especially when computing resources are limited. Summary of the invention

[0011] In order to overcome the shortcomings of the prior art, the present invention provides a method for identifying key parts using regional perception, which aims to achieve accurate identification of key parts of the target through efficient regional perception and adaptive attention mechanism, and is particularly suitable for environments with limited computing resources such as embedded systems. On the basis of preliminarily completing target recognition, the present invention uses the obtained perception information to directly allocate detailed attention to the target area instead of the entire image, significantly reducing the computational complexity and improving the processing efficiency. Through regional perception technology, the present invention can intelligently focus on the target area and avoid redundant processing of background information. Subsequently, the adaptive attention module dynamically adjusts its work focus, performs in-depth analysis of key parts in the target area, effectively captures and distinguishes the characteristics of these parts, thereby achieving high-precision key part recognition. This method not only reduces unnecessary computational overhead, but also ensures the accuracy and robustness of key part recognition. By optimizing the computational process and intelligent resource allocation, the present invention provides an efficient target and key part recognition solution for embedded systems and other resource-constrained scenarios, which is expected to improve the intelligent level of intelligent monitoring, automated driving, medical image analysis and other fields, and promote the intelligent development and application of the industry.

[0012] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0013] Step 1: Target detection and region preselection;

[0014] Use the target detection algorithm to perform global detection on the input image to locate all possible targets;

[0015] Output the bounding box coordinates of the target;

[0016] Step 2: Region perception and feature extraction;

[0017] Based on the results of target detection, focus on the area where each target is located; use the feature map generated during the detection network reasoning, combined with the detection results in step 1, to perform feature extraction and information fusion;

[0018] Step 3: Adaptive attention allocation;

[0019] Use adaptive attention mechanism to dynamically adjust attention distribution so that the key part recognition model focuses on the key parts of the target rather than the background or secondary areas;

[0020] Step 4: Identification and analysis of key parts;

[0021] Use the features filtered by the attention mechanism to identify and analyze key areas;

[0022] Step 5: Result output and optimization feedback;

[0023] The identified key parts information is integrated to form the final recognition result. At the same time, data from the execution process is collected for continuous optimization and parameter adjustment of the model.

[0024] Furthermore, the step 1 is specifically as follows:

[0025] Step 1-1: After obtaining the original or image-adjusted image from the image information source, a deep convolutional neural network is used to perform multi-layer convolution and pooling operations on the input image to extract feature maps at different levels, automatically learn the feature patterns related to the target in the image, and thus capture the target feature information;

[0026] Step 1-2: Based on the prior knowledge of the target shape and size, by presetting a variety of different anchor boxes, the possibility of detecting targets of different sizes and shapes is increased, thereby improving the recall rate of detection;

[0027] Step 1-3: Use the non-maximum suppression algorithm to suppress bounding boxes with high overlap and low scores to ensure the accuracy of the final output bounding box.

[0028] Furthermore, the step 2 is specifically as follows:

[0029] Step 2-1: Based on the target area bounding box determined by target detection, combined with the feature map generated during the detection network inference, the corresponding area is processed using a lightweight feature extraction module;

[0030] Step 2-2: Through weighted summation, concatenation, and the use of attention mechanisms, the features at different levels are fused according to certain rules to obtain a more comprehensive and representative feature representation;

[0031] Step 2-3: Share feature information with the detection network, realize regional perception based on the detection results, and retain key features for subsequent identification of key parts of the target.

[0032] Furthermore, the step 3 is specifically as follows:

[0033] Step 3-1: Introduce a learnable weight matrix, which has the same dimension as the feature map. For the eigenvalue at each position, calculate the product of the element at the corresponding position in the weight matrix to obtain the weighted eigenvalue;

[0034] Assume the input feature map is F∈R H×W×C , where H and W represent the height and width of the feature map respectively, C represents the number of channels, and the learnable weight matrix is ​​A∈R H×W×C ;

[0035] The calculation process of attention allocation is expressed as: F attended =F⊙A, where ⊙ represents element-wise multiplication operation.

[0036] Step 3-2: Normalize the weighted feature map.

[0037] Furthermore, the step 5 is specifically as follows:

[0038] Step 5-1: Arrange the key parts information into a unified data structure, and perform data verification and exception processing before output;

[0039] Step 5-2: Collect various data during execution, including input, intermediate and final results, system running time and resource usage, and store them in the database;

[0040] Step 5-3: Establish accuracy, recall, F1 value, and average precision to evaluate recognition accuracy, use processing time to measure processing speed, and use memory and computing resource usage to evaluate resource consumption; based on data and indicators, adjust model parameters such as confidence.

[0041] The beneficial effects of the present invention are as follows:

[0042] Compared with the method of realizing key part detection by simple target detection and segmentation, the present invention can greatly reduce the amount of calculation and the labeling workload of the preliminary tasks under the condition of meeting the actual accuracy requirements, reduce the algorithm's requirements for hardware and data volume, and thus reduce the cost of actual deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the system framework and flow chart of the present invention;

[0044] Figure 2 This is a diagram introducing the overall network architecture and functions of the present invention;

[0045] Figure 3 is an example of the result diagram of the region perception and attention allocation process of the present invention, (a) coarse granularity, (b) standard granularity, (c) high granularity;

[0046] Figure 4 is a schematic diagram of training results of an embodiment of the present invention;

[0047] Figure 5 The following are examples of experimental results of embodiments of the present invention: (a) rail and obstacle recognition, (b) ship, island and hull recognition, and (c) tank and turret track recognition. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] The present invention aims to solve the contradiction between computational efficiency and resource limitation in target and key part recognition in embedded systems, as well as the difficulty of maintaining high-precision recognition in complex scenarios. By introducing regional perception and adaptive attention mechanism, efficient and accurate target key part recognition is achieved, which is particularly suitable for resource-constrained environments and reduces system deployment costs.

[0050] The present invention achieves accurate recognition of key parts of the target by introducing efficient region perception technology and adaptive attention mechanism. The specific objectives are as follows:

[0051] (1) Improved computing efficiency: By intelligently focusing on the target area rather than the entire image and directly utilizing the detection network feature information, the computational complexity is significantly reduced and processing efficiency is improved, making it suitable for environments with limited computing resources, such as embedded systems.

[0052] (2) Enhanced robustness: Even in complex scenarios, such as those with changing lighting or cluttered backgrounds, it is able to maintain high-precision recognition of key target parts and reduce the impact of external factors on the recognition effect.

[0053] (3) Optimize resource allocation: Use the adaptive attention module to perceive the key parts of the target and use feature information for regional perception rather than pixel-level analysis to avoid redundant processing of background information and reduce unnecessary computational overhead.

[0054] In summary, the objective of the present invention is to provide an efficient and accurate solution for identifying targets and their key parts, which is particularly suitable for resource-constrained environments, so as to address the deficiencies of the prior art in terms of computing efficiency, real-time processing capability, sensitivity to external factors, and cost control.

[0055] The steps of the method of the present invention are as follows:

[0056] Step 1: Target detection and region preselection;

[0057] First, the object detection algorithm is used to perform global detection on the input image to locate all possible objects. This step outputs the bounding box coordinates of the object as the basis for subsequent region perception.

[0058] Objective: To ensure that the initial positioning of the target is accurate and provide an effective regional scope for subsequent detailed analysis.

[0059] Technical highlights: Utilize the feature extraction capability of the deep learning model, combined with the anchor box mechanism and non-maximum suppression algorithm to obtain a high-precision target bounding box.

[0060] Step 2: Region perception and feature extraction;

[0061] Based on the results of target detection, we intelligently focus on the area where each target is located to avoid unnecessary calculations on the entire image. We use the feature map generated during the detection network reasoning and combine it with the detection results in step 1 to perform further feature extraction and information fusion.

[0062] Objective: By directly using the feature maps generated by the inference of the detection network, the amount of calculation can be greatly reduced and the processing speed can be improved, while retaining enough information for the identification of key parts.

[0063] Technical highlights: A lightweight feature extraction module is used to directly obtain features from the detection network in step one, reducing redundant calculations and optimizing resource utilization.

[0064] Step 3: Adaptive attention allocation;

[0065] Description: Based on the extracted feature information, the adaptive attention mechanism is used to dynamically adjust the attention distribution, so that the key part recognition model pays more attention to the key parts of the target rather than the background or secondary areas.

[0066] Objective: Ensure that computing resources are allocated preferentially to areas with the most information value, and improve the accuracy and efficiency of key area identification.

[0067] Technical highlights: Through a learnable weight matrix, attention weights are automatically allocated according to the statistical characteristics in the feature map to focus on key areas.

[0068] Step 4: Identification and analysis of key parts;

[0069] Description: Use the features filtered by the attention mechanism to identify and analyze key areas.

[0070] Objective: Maintain or improve the accuracy and robustness of key part recognition while reducing computing resource consumption.

[0071] Technical highlights: Combine the predictive ability of deep learning models and the guiding role of attention mechanisms to extract features, accurately locate and classify key parts.

[0072] Step 5: Result output and optimization feedback;

[0073] Description: Integrate the identified key parts information to form the final recognition result, and collect data during the execution process for continuous optimization and parameter adjustment of the model.

[0074] Objective: To provide accurate and timely key part identification results, and gradually improve the performance and adaptability of the system through a feedback mechanism.

[0075] Technical points: Establish model evaluation indicators, monitor recognition accuracy, processing speed and resource consumption, and adjust algorithm parameters in a timely manner to meet the needs of different scenarios.

[0076] The specific measures and functions of each step are as follows:

[0077] Target detection and region preselection;

[0078] After obtaining the original or image-adjusted image from the image information source, a deep convolutional neural network (CNN) is used to perform multi-layer convolution and pooling operations on the input image to extract feature maps at different levels, automatically learn the feature patterns related to the target in the image, and effectively capture the target feature information. Based on the prior knowledge of the shape and size of the target, by presetting a variety of different anchor boxes, the possibility of detecting targets of different sizes and shapes is increased, and the recall rate of detection is improved. The non-maximum suppression algorithm is used based on the need to reduce repeated detection and improve detection accuracy. By suppressing bounding boxes with high overlap and low scores, the bounding boxes of the final output are ensured to be more representative and accurate. Based on the above means, accurate preliminary positioning of the target in the input image is achieved, providing an effective area range for subsequent fine analysis, and improving the performance and accuracy of target detection and area preselection.

[0079] Region perception and feature extraction;

[0080] In the regional perception stage, by focusing only on the area where the target is located, the entire image is avoided from being fully processed, thus significantly reducing the amount of computation. Because most background areas are often irrelevant for the identification and analysis of specific targets, focusing on key areas can improve computational efficiency. Based on the target area bounding box determined by target detection, combined with the feature map generated during the inference of the detection network, the corresponding area is processed using a specially designed lightweight feature extraction module. These feature maps are obtained through network learning and extraction, and have certain semantic and spatial information, which can serve as the basis for subsequent feature extraction and analysis. Through weighted summation, splicing, and the use of attention mechanisms, features at different levels are fused according to certain rules to obtain a more comprehensive and representative feature representation. By sharing feature information with the detection network and realizing regional perception based on the detection results, it is possible to reduce the amount of computation and improve the processing speed during regional perception and feature extraction, while retaining key features for subsequent target key part identification.

[0081] Adaptive attention allocation;

[0082] The extracted feature information is analyzed and processed, and the feature information usually exists in the form of a multi-dimensional feature map. By introducing a learnable weight matrix, which has the same dimension as the feature map, for the feature value at each position, the product of it and the element at the corresponding position in the weight matrix is ​​calculated to obtain the weighted feature value. Then, the weighted feature map is normalized. The principle of the adaptive attention mechanism is to dynamically allocate computing resources according to the importance of the input features. By learning and identifying the key patterns and information distribution in the feature map, the model can automatically focus more attention on the areas that are most valuable for key part identification. This mechanism can effectively capture local and global dependencies in the feature map, highlight important features, and suppress the influence of irrelevant or minor features, thereby improving the performance and efficiency of the model.

[0083] Assume that the input feature map is F∈R H×W×C , where H and W represent the height and width of the feature map respectively, and C represents the number of channels. The learnable weight matrix is ​​A∈R H×W×C .

[0084] The calculation process of attention allocation can be expressed as: F attended =F⊙A, where ⊙ represents the element-level multiplication operation. Then normalization is performed. By continuously optimizing the weight matrix A, the model can learn the optimal attention allocation strategy during the training process, thereby achieving adaptive attention allocation, focusing more computing resources on key parts, and improving the accuracy and efficiency of key part recognition.

[0085] Identification and analysis of key parts;

[0086] With the help of processed features, the perception results and identification of key parts are obtained. For the location and classification of key parts, a combination of regression analysis and classification algorithms is used. Regression analysis is used to predict the position and shape of key parts, while the classification algorithm determines the category to which they belong. The powerful representation and feature learning capabilities of deep learning models are fully utilized, combined with the guidance of the attention mechanism, to accurately identify and locate key parts from complex input features.

[0087] Result output and optimization feedback;

[0088] When integrating and outputting results, organize key information into a unified data structure, present it in a clear way, and perform data verification and exception handling before output. Collect various types of data during execution, including input, intermediate and final results, system running time and resource usage, and store them efficiently in the database. Establish accuracy, recall rate, F1 value, average precision mean, etc. to evaluate recognition accuracy, use processing time to measure processing speed, and use memory and computing resource usage to evaluate resource consumption. Based on data and indicators, adjust model parameters such as confidence.

[0089] Figure 2 Introducing the overall network architecture and functions

[0090] The method of the present invention can reduce the requirements for hardware and deployment costs, and can more efficiently obtain information on key parts of the target compared to simple target detection.

[0091] Example:

[0092] In order to verify the effectiveness of the algorithm in this paper, the results of feature information utilization and detection algorithm are shown below, as well as the actual implementation effect diagram in the final practical application process. According to actual needs, the algorithms in each link can be adjusted to enable it to have multiple capabilities such as target detection, key part detection, and parameter adaptation.

[0093] Illustration of the results of the region perception and attention allocation process. Figure 3 As shown:

[0094] Figure 3 As a result diagram of the process of using different fine-grained regional perception and attention allocation for different scenarios in a target-free scene, for example, coarse and fine-grained regional perception and attention allocation are used in the remote sensing field to focus on non-water areas, standard fine-grained regional perception and attention allocation are used in the aerial photography field to focus on the differences in characteristics of different regions, and high-grained regional perception and attention allocation are used in the road traffic field to ensure safety, focusing on suspected obstacle areas.

[0095] The present invention uses mAP (mean Average Precision) as the evaluation index for detection. mAP is the average value of AP of each category, which can reflect the average precision of the model; AP is the area under the PR curve, which is used to represent the average precision of the category. The PR curve is the Precision-Recall curve, where the calculation formulas of Precision and Recall are as follows:

[0096]

[0097] The above network is iteratively trained using the test data set. The model results of each round of training are saved, and the loss value is visualized. After the model training is completed, the test set is used for testing to obtain the mAP value of the model. The corresponding loss value, mAP, Precision, Recall, and Precision-Recall curve of the training are as follows. Figure 4 shown.

[0098] The experimental results of the method of the present invention are exemplified as follows: Figure 5 , presenting the recognition results of rails and obstacles, ships, islands and hulls, and tanks, turrets and tracks. In the actual deployment process, the corresponding target data set can be selected for training according to needs, and the algorithm can be slightly adjusted according to the scenario to give full play to the generalization ability of deep learning technology and reduce deployment costs.

[0099] The method of the present invention achieves efficient and accurate recognition of the key parts of the target through the above steps. From target detection to regional perception, attention allocation, and then to key part recognition and analysis, each step is aimed at optimizing the use of computing resources and improving the accuracy and robustness of recognition, which is particularly suitable for resource-constrained environments such as embedded systems. Through continuous data feedback and model optimization, the method of the present invention can continuously improve its performance in complex scenarios, providing strong technical support for the fields of intelligent monitoring, automated driving, and medical image analysis.

Claims

1. A method for identifying key parts using region perception, characterized in that: The steps include: Step 1: Target detection and region preselection; Use the target detection algorithm to perform global detection on the input image to locate all possible targets; Output the bounding box coordinates of the target; Step 2: Region perception and feature extraction; Based on the results of target detection, focus on the area where each target is located; use the feature map generated during the detection network reasoning, combined with the detection results in step 1, to perform feature extraction and information fusion; Step 3: Adaptive attention allocation; Use adaptive attention mechanism to dynamically adjust attention distribution so that the key part recognition model focuses on the key parts of the target rather than the background or secondary areas; Step 4: Identification and analysis of key parts; Use the features filtered by the attention mechanism to identify and analyze key areas; Step 5: Result output and optimization feedback; The identified key parts information is integrated to form the final recognition result. At the same time, data from the execution process is collected for continuous optimization and parameter adjustment of the model.

2. The method for identifying key parts using region perception according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1-1: After obtaining the original or image-adjusted image from the image information source, a deep convolutional neural network is used to perform multi-layer convolution and pooling operations on the input image to extract feature maps at different levels, automatically learn the feature patterns related to the target in the image, and thus capture the target feature information; Step 1-2: Based on the prior knowledge of the target shape and size, by presetting a variety of different anchor boxes, the possibility of detecting targets of different sizes and shapes is increased, thereby improving the recall rate of detection; Step 1-3: Use the non-maximum suppression algorithm to suppress bounding boxes with high overlap and low scores to ensure the accuracy of the final output bounding box.

3. The method for identifying key parts using region perception according to claim 2, characterized in that: The step 2 is specifically as follows: Step 2-1: Based on the target area bounding box determined by target detection, combined with the feature map generated during the detection network inference, the corresponding area is processed using a lightweight feature extraction module; Step 2-2: Through weighted summation, concatenation, and the use of attention mechanisms, the features at different levels are fused according to certain rules to obtain a more comprehensive and representative feature representation; Step 2-3: Share feature information with the detection network, realize regional perception based on the detection results, and retain key features for subsequent identification of key parts of the target.

4. The method for identifying key parts using region perception according to claim 3, characterized in that: The step 3 is specifically as follows: Step 3-1: Introduce a learnable weight matrix, which has the same dimension as the feature map. For the eigenvalue at each position, calculate the product of the element at the corresponding position in the weight matrix to obtain the weighted eigenvalue; Assume the input feature map is F∈R H×W×C , where H and W represent the height and width of the feature map respectively, C represents the number of channels, and the learnable weight matrix is ​​A∈R H×W×C ; The calculation process of attention allocation is expressed as: F attended =F⊙A, where ⊙ represents the element-wise multiplication operation; Step 3-2: Normalize the weighted feature map.

5. The method for identifying key parts using region perception according to claim 4, characterized in that: The step 5 is specifically as follows: Step 5-1: Arrange the key parts information into a unified data structure, and perform data verification and exception processing before output; Step 5-2: Collect various data during execution, including input, intermediate and final results, system running time and resource usage, and store them in the database; Step 5-3: Establish accuracy, recall, F1 value, and average precision to evaluate recognition accuracy, use processing time to measure processing speed, and use memory and computing resource usage to evaluate resource consumption; Based on data and indicators, adjust model parameters such as confidence.

Images