Aerial image small target detection algorithm based on Focus-and-Detect framework
By applying the detection algorithm of the Focus-and-Detect framework in aerial images, combined with the incomplete frame suppression method, the problem of poor detection performance of small targets in the prior art is solved, and the small target detection effect with high accuracy, efficiency and robustness is achieved.
Patent Information
- Application Number
- CN202411920624.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-25
- Publication Date
- 2025-04-25
AI Technical Summary
The existing deep learning models have poor performance in small and medium-sized aerial images and complex scenarios, with high computational complexity, low recognition accuracy and poor generalization ability.
The aerial image small object detection algorithm based on the Focus-and-Detect framework is adopted, and the precise detection of small objects in aerial images is achieved through the two stages of focus and detection combined with the incomplete frame suppression method. The algorithm includes steps such as image preprocessing, focus area generation, fine detection and incomplete frame suppression.
It significantly improves the detection accuracy of small targets in aerial images, overcomes the truncation effect, improves the integrity and execution efficiency of target detection, has good robustness and generalization capabilities, and is suitable for a variety of scenarios and complex environments.
Smart Images
Figure CN119888484A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of target detection algorithms, and in particular relates to an aerial image small target detection algorithm based on a Focus-and-Detect framework. Background Art
[0002] In the field of image processing, especially in object detection, deep learning and convolutional neural networks (CNN) are the current mainstream technologies. These technologies have achieved remarkable results in image recognition, object detection, etc. For example, convolutional neural networks have performed well in multiple tasks such as image classification and object detection due to their powerful feature extraction capabilities.
[0003] However, despite the great progress made by deep learning technology in object detection, some key issues and limitations still exist. First, high computational complexity is a major problem. Deep learning models usually require a lot of computing resources and time to train, especially when dealing with large-scale datasets, which limits their use in real-time applications. Second, low recognition accuracy is also a challenge. Although deep learning models can achieve near-human-level accuracy in some cases, their performance is still not as good as expected in complex scenarios, such as small object detection or occluded object detection.
[0004] In addition, poor generalization ability is also an important shortcoming of existing technologies. Deep learning models often rely on a large amount of labeled data for training, which makes them perform poorly when faced with unseen data or different scenarios. Summary of the invention
[0005] In view of the technical problem that the performance of the above-mentioned deep learning model is still not as good as expected in complex scenes, such as small target detection or occluded target detection, the present invention provides an aerial image small target detection algorithm based on the Focus-and-Detect framework, aiming to solve the problem of poor performance of aerial image small target detection in the prior art. The algorithm achieves accurate detection of small targets in aerial images through the two stages of focusing and detection, combined with the incomplete frame suppression method.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0007] The small target detection algorithm for aerial images based on the Focus-and-Detect framework includes the following steps:
[0008] S1, receiving the aerial image as the input of the algorithm, uniformly adjusting the input aerial image to a preset size, and using the filter image processing technology to remove the noise in the image and improve the image quality;
[0009] S2, generating a focus area;
[0010] S3, target detection in the focus area;
[0011] S4, perform incomplete frame suppression;
[0012] S5. Summarize the targets detected in the second stage, including the location and category information of the targets, and output the summarized detection results in a preset format for use in subsequent tasks.
[0013] The method for removing noise from the image in S1 is: combining the median pre-filtering and wavelet transform methods to denoise the processed image, thereby better retaining the image edge while removing the noise.
[0014] The method for generating the focus area in S2 is:
[0015] S2.1. Gaussian mixture model modeling: Gaussian mixture model GMM is used to model the targets in the aerial image. The Gaussian mixture model GMM captures the information of the shape, size and distribution of the target, and the parameters of the model are obtained through training;
[0016] S2.2. Object detector network construction: Design a convolutional neural network as an object detector. The network uses a Gaussian mixture model (GMM) as a supervisory signal. The network structure includes a feature extraction layer, a classification layer, and a regression layer for preliminary detection of objects in the image.
[0017] S2.3, Target cluster generation: Based on the output results of the object detector network, a clustering algorithm is used to divide the detected targets into different clusters. These clusters represent the areas that may contain small targets, namely the focus areas.
[0018] The method for detecting the target within the focus area in S3 is:
[0019] S3.1. Fine detection network construction: In the focus area, another more fine object detector network is constructed. The network structure is more complex and contains more convolutional layers, pooling layers, and fully connected layers to improve detection accuracy.
[0020] S3.2, Target recognition and positioning: Use the fine detection network to more accurately identify and locate the target in the focus area. By extracting the feature information of the target, the classification layer determines the category of the target, and the regression layer predicts the location of the target.
[0021] The method for performing incomplete frame suppression in S4 is:
[0022] S4.1, Truncation effect analysis: In the area search method, due to the limitation of the search window, it will cause the truncation effect of the target, that is, part of the target is truncated by the window;
[0023] S4.2, incomplete box suppression: by analyzing the relationship between the detection box and the image boundary, determine whether there is a truncation effect. For target boxes that may be truncated, adopt a suppression or correction strategy, adjust the size or position of the box, or merge it with other boxes to improve the completeness of target detection.
[0024] Compared with the prior art, the present invention has the following beneficial effects:
[0025] The present invention can accurately identify and locate small targets in aerial images and significantly improve detection accuracy through the two-stage detection strategy of the Focus-and-Detect framework. The application of the IBS method effectively overcomes the truncation effect, improves the integrity of target detection, and further improves the detection performance. By optimizing the network structure and introducing acceleration technology, the execution efficiency of the algorithm is significantly improved in the present invention, which can meet the needs of real-time applications. The high efficiency of the algorithm enables it to have better scalability and practicality when processing large-scale aerial image data. Since the present invention adopts a Gaussian mixture model to model the target, the algorithm has stronger adaptability to changes in the shape, size and distribution of the target. Tests of the algorithm in various scenarios and complex environments show that it has good robustness and generalization ability and can cope with various challenges. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.
[0027] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.
[0028] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than to limit the claims of the present invention. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.
[0030] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0031] Small target detection algorithm for aerial images based on the Focus-and-Detect framework, such as Figure 1 As shown, the following steps are included:
[0032] Step 1: Receive the aerial image as the input of the algorithm, adjust the input aerial image to the preset size, use the filter image processing technology to remove the noise in the image and improve the image quality. Combine the median pre-filtering and wavelet transform methods to denoise the processed image, which can better retain the image edge while removing the noise.
[0033] Step 2: Generate a focus area.
[0034] Step 2.1: Gaussian mixture model modeling: Gaussian mixture model GMM is used to model the targets in the aerial image. The Gaussian mixture model GMM captures the information of the shape, size and distribution of the target, and the parameters of the model are obtained through training.
[0035] Step 2.2, object detector network construction: Design a convolutional neural network as an object detector. The network uses a Gaussian mixture model GMM as a supervisory signal. The network structure includes a feature extraction layer, a classification layer, and a regression layer for preliminary detection of objects in the image.
[0036] Step 2.3, target cluster generation: Based on the output results of the object detector network, a clustering algorithm is used to divide the detected targets into different clusters. These clusters represent the areas that may contain small targets, i.e., the focus areas.
[0037] Step 3: Target detection within the focus area.
[0038] Step 3.1: Build a refined detection network: In the focused area, build another more refined object detector network with a more complex structure, including more convolutional layers, pooling layers, and fully connected layers to improve detection accuracy.
[0039] Step 3.2, target recognition and positioning: Use the fine detection network to more accurately identify and locate the target in the focus area. By extracting the feature information of the target, the classification layer determines the category of the target, and the regression layer predicts the location of the target.
[0040] Step 4: Perform incomplete frame suppression.
[0041] Step 4.1, truncation effect analysis: In the area search method, due to the limitation of the search window, it will cause the truncation effect of the target, that is, part of the target is truncated by the window.
[0042] Step 4.2, incomplete box suppression: By analyzing the relationship between the detection box and the image boundary, determine whether there is a truncation effect. For target boxes that may be truncated, adopt a suppression or correction strategy, adjust the size or position of the box, or merge it with other boxes to improve the integrity of target detection.
[0043] Step 5: Summarize the targets detected in the second stage, including the location and category information of the targets, and output the summarized detection results in a preset format for use in subsequent tasks.
[0044] Through the above steps, the aerial image small target detection algorithm based on the Focus-and-Detect framework proposed in the present invention can achieve accurate detection of small targets in aerial images. The algorithm combines the two stages of focus and detection, as well as the incomplete frame suppression method, effectively improving the accuracy and efficiency of target detection, and providing strong support for subsequent tasks such as target recognition and scene understanding of aerial images.
[0045] In order to verify the effectiveness and superiority of the aerial image small target detection algorithm based on the Focus-and-Detect framework proposed in this invention, this embodiment uses the VisDrone validation dataset as the experimental object for testing and analysis. The VisDrone dataset is a large-scale aerial image dataset that contains rich scenes and diverse target types, and is very suitable for evaluating the performance of small target detection algorithms.
[0046] During the experiment, the present embodiment firstly preprocessed the data set, including size normalization and denoising operations, to improve the image quality and reduce the influence of noise on the detection results. Then, the present embodiment conducted experiments according to the algorithm flow of the present invention.
[0047] In the first stage of the Focus-and-Detect framework, this embodiment uses a Gaussian mixture model (GMM) to model the objects in the VisDrone dataset, and obtains a GMM model suitable for the dataset through training. Next, this embodiment constructs an object detector network and trains it using the GMM model as a supervisory signal. The network can generate preliminary target clusters based on the output of the GMM model to form a focus area.
[0048] In the second stage, the embodiment performs refined target detection in the focus area. The embodiment builds a more complex and refined object detector network, which achieves accurate recognition and positioning of the target by extracting deep features of the target and performing classification and regression operations.
[0049] In addition, this embodiment also implements an incomplete box suppression (IBS) method to overcome the truncation effect that may exist in the area search method. By analyzing the relationship between the detection box and the image boundary, this embodiment suppresses or corrects the target box that may be truncated, thereby improving the integrity of target detection.
[0050] Finally, this embodiment summarizes the targets detected in the second stage and outputs the final detection results. In order to evaluate the performance of the algorithm, this embodiment uses the average precision (AP) as the evaluation indicator. By adjusting the network parameters and training strategy, this embodiment achieves an AP score of 42.06 on the VisDrone validation dataset.
[0051] Compared with other small target detection methods reported in existing literature, the algorithm of the present invention shows obvious advantages in AP score, which fully proves the effectiveness and superiority of the algorithm of the present invention in the task of small target detection in aerial images.
[0052] In summary, the aerial image small target detection algorithm based on the Focus-and-Detect framework proposed in the present invention achieves accurate detection of small targets in aerial images by focusing first and then detecting, combined with the incomplete frame suppression method. The algorithm has high accuracy and efficiency, and provides strong support for subsequent tasks such as target recognition and scene understanding of aerial images. Through experimental verification on the VisDrone verification dataset, this embodiment further proves the feasibility and superiority of the algorithm in practical applications.
[0053] Only the preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the protection scope of the present invention.
Claims
1. Aerial image small target detection algorithm based on the Focus-and-Detect framework, characterized by: The following steps are involved: S1, receiving the aerial image as the input of the algorithm, uniformly adjusting the input aerial image to a preset size, and using the filter image processing technology to remove the noise in the image and improve the image quality; S2, generating a focus area; S3, target detection in the focus area; S4, perform incomplete frame suppression; S5. Summarize the targets detected in the second stage, including the location and category information of the targets, and output the summarized detection results in a preset format for use in subsequent tasks.
2. The small target detection algorithm for aerial images based on the Focus-and-Detect framework according to claim 1, characterized in that: The method for removing noise from the image in S1 is: combining the median pre-filtering and wavelet transform methods to denoise the processed image, thereby better retaining the image edge while removing the noise.
3. The small target detection algorithm for aerial images based on the Focus-and-Detect framework according to claim 1, characterized in that: The method for generating the focus area in S2 is: S2.
1. Gaussian mixture model modeling: Gaussian mixture model GMM is used to model the targets in the aerial image. The Gaussian mixture model GMM captures the information of the shape, size and distribution of the target, and the parameters of the model are obtained through training; S2.
2. Object detector network construction: Design a convolutional neural network as an object detector. The network uses a Gaussian mixture model (GMM) as a supervisory signal. The network structure includes a feature extraction layer, a classification layer, and a regression layer for preliminary detection of objects in the image. S2.3, Target cluster generation: Based on the output results of the object detector network, a clustering algorithm is used to divide the detected targets into different clusters. These clusters represent the areas that may contain small targets, namely the focus areas.
4. The small target detection algorithm for aerial images based on the Focus-and-Detect framework according to claim 1, characterized in that: The method for detecting the target within the focus area in S3 is: S3.
1. Fine detection network construction: In the focus area, another more fine object detector network is constructed. The network structure is more complex and contains more convolutional layers, pooling layers, and fully connected layers to improve detection accuracy. S3.2, Target recognition and positioning: Use the fine detection network to more accurately identify and locate the target in the focus area. By extracting the feature information of the target, the classification layer determines the category of the target, and the regression layer predicts the location of the target.
5. The small target detection algorithm for aerial images based on the Focus-and-Detect framework according to claim 1, characterized in that: The method for performing incomplete frame suppression in S4 is: S4.1, Truncation effect analysis: In the area search method, due to the limitation of the search window, it will cause the truncation effect of the target, that is, part of the target is truncated by the window; S4.2, incomplete box suppression: by analyzing the relationship between the detection box and the image boundary, determine whether there is a truncation effect. For target boxes that may be truncated, adopt a suppression or correction strategy, adjust the size or position of the box, or merge it with other boxes to improve the completeness of target detection.