A SINGLE-SHOT OBJECT DETECTION METHOD

The Siamese convolutional network-based method addresses the challenges of labeled data scarcity and object orientation in remote sensing by generating orientation-independent candidate boxes for accurate single-shot object detection with reduced costs.

TR202305259A3Active Publication Date: 2026-10-21HAVELSAN HAVA ELEKTRONIK SANAYI VE TICARET ANONIM SIRKETI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
TR202305259
Authority / Receiving Office
TR · TR
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-11
Publication Date
2026-10-21

AI Technical Summary

Technical Problem

Existing single-shot object detection methods in remote sensing are hindered by the lack of labeled data, high cost of data labeling, and the negative impact of object orientation on performance, particularly in scenarios with limited data access.

Method used

A single-shot object detection method using a Siamese convolutional network that processes query and support images separately, generating orientation-independent candidate boxes through feature extraction and fusion, followed by classification and regression to accurately detect objects.

Benefits of technology

Enables accurate and efficient object detection with reduced hardware and labeling costs, independent of object orientation, even in scenarios with limited data access.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000011_0000
    Figure 00000011_0000
  • Figure 00000012_0000
    Figure 00000012_0000
  • Figure 00000012_0001
    Figure 00000012_0001
Patent Text Reader

Abstract

This relates to a one-shot object detection method (44) which involves processing a query image tensor (19) with candidate box templates (20) in accordance with the provision of the feature tensor (18), applying a directional candidate box extraction operation (22) to generate candidate boxes and subsequently applying a direction-independent candidate region of interest finding operation (26) to the query image attributes (24) that generate the directional candidate boxes.
Need to check novelty before this filing date? Find Prior Art

Description

1 TARIFF A SINGLE-SHOT OBJECT DETECTION METHOD Technical Area The invention involves creating an attribute tensor such that the query and supporting images correspond separately. It relates to a single-shot object detection method that includes providing 5. Previous Technique A study on object detection using supervised methods in the field of remote sensing in the literature. These methods are dependent on labeled data and easy access to labeled data. The absence of objects in remote sensing images creates difficulties. The difference in their orientations increases the cost of data labeling. Also, remotely 10 The fact that objects are oriented in the sensing field negatively affects the performance of single-shot methods. It has an impact as follows. In the technical field, artificial neural network-based solutions have been identified for distant target detection. These include... CN113408546 is a single mechanism based on reciprocal global context attention. This relates to a sample target identification method. The invention describes the characteristics of the input query image and 15 It creates a feature extraction module to obtain the characteristics of the supporting image; Obtaining the general context property of the query image based on the query image's property. Create a global context module for and support the image based on its properties. The general context feature of the image is obtained; the query image at the channel level. combining addiction information and supporting image features and supporting 20 channel-level dependency information of the image and characteristics of the query image. A fusion module was created to combine; from the merged query image A region suggestion module is created to obtain the region of interest; category independent. Create a classifier, identify the characteristics of the region of interest, and combine support. Combine the features of the image and the 25 of interest through classification features. determining whether the region is an area that is a target category and the output model, query The reciprocal global context, which estimates the position of the target category in the image, is noteworthy. It describes the solution related to a single-example target detection method based on its mechanism. However, the candidate does not involve the application of a box extraction (finding) process. Furthermore, the candidate is of interest. The application of the region finding process is not available. 30 2 In conclusion, all the problems mentioned above necessitate an innovation in the relevant technical field. It has made it mandatory. Purposes of the Invention The present invention aims to eliminate the aforementioned disadvantages and contribute to the relevant technical field. It relates to a single-shot object detection method designed to bring new advantages. 5 One aim of the invention is to determine the possible locations of target objects on a query image. using pre-cropped support images of the class to be identified In remote sensing images where this was made possible, a single orientation-independent image was detected. It is related to projectile object detection methods. Another purpose of the invention is to transmit orientation-differentiated data containing single-shot object position and class information. It is a single-system tool that can operate in clusters and is used in areas where data access is difficult. It is related to projectile object detection methods. Brief Description of the Invention All the purposes mentioned above and those that will emerge from the detailed explanation below. To achieve this, the existing invention, inquiry and supporting images are oriented from 15 It entered a common Siamese convolutional network that was unaffected and enabled feature extraction. The query and support images are processed separately by passing them through a feature extraction network. A single-shot object detection system that involves providing a corresponding feature tensor. It is a method. The invention provides a query image suitable for providing an attribute tensor. 20 oriented towards the production of candidate boxes, where the tensor is processed with candidate box templates. The candidate box extraction process is applied, and the query generates approximately oriented candidate boxes. the process of finding a candidate region of interest independent of an orientation to image attributes by obtaining a single-shot object detection result suitable for implementation This characterizes the singular orientation of objects in query-support images. This ensures accurate location and positioning of projectiles through projectile detection. 25 A key feature of a possible application of the invention is that the feature tensor of the support images can also be included. The traditional candidate interest area identification process involves implementing the application step. Thus, feature extractors extract instance-level features that are unaffected by orientation. is being removed. 3 A key feature of a possible application of the invention is the generation of attributes at the sample level. candidate interest region and traditional candidate interest region procedures, independent of orientation. A query involves the process of obtaining a candidate region of interest result. Thus, the query... By obtaining sample-level results from the images, the merging step begins. It can be transferred. 5 A key feature of a possible application of the invention is the generation of attributes at the sample level. candidate interest region and traditional candidate interest region procedures, independent of orientation. This involves the process of obtaining a support candidate region of interest result. Thus The merging step involves obtaining sample-level results from supporting images. It can be transferred. 10 A characteristic of a possible application of the invention is a query candidate region of interest and a support candidate. This involves the process of combining the results corresponding to the region of interest. Thus The result obtained by combining the query candidate interest region result with the support candidate interest region result. Classification and regression can be performed on the obtained data. The characteristic of a possible application of the invention is the example 15 provided by combining the results. The process involves classifying and regressing based on attributes at the level using object detection. This involves a step that allows attributes to be obtained at the sample level. A key feature of a possible application of the invention is obtaining a single-shot object detection result. This is achieved by using data obtained through classification and regression analysis. Thus, object detection is performed using features obtained at the sample level. 20 Brief Description of the Figures Figure 1 shows the flowchart of the method relating to a single-shot object detection method, which is the subject of the invention. The illustration is given. Figure 2 shows a study of a single-shot object detection method, which is the subject of the invention. A flowchart representation of the scenario is given. 25 Figure 3 shows a working scenario of a single-shot object detection method given in Figure 2. A schematic representation of the candidate box distributions is given. 4 Detailed Description of the Invention This detailed explanation describes a single-shot object detection method, which is the subject of the invention, but only in a broader sense. This is explained with examples that do not create any limiting effects on a better understanding. Figure 1 shows the flowchart of the method relating to a single-shot object detection method, which is the subject of the invention. It is shown. In a one-shot object detection method, firstly, query images (11) and 5 common supporting images (12) that are not affected by orientation and enable feature extraction. Input is given to a Siamese convolutional network (13) (14). Then from the feature extraction network by passing through (16) the interrogation and support images (11) (12) separately to correspond The feature tensor is provided (18). In the invention, the single-shot object detection method (10) then a query image in accordance with the provision of the feature tensor (18) 10 The tensor (19) was processed with candidate box templates (20), and the candidate boxes were produced. A directional candidate box extraction / finding process is applied (22). Here, the candidate box The templates (20) are candidate box templates with different sizes, aspect ratios and orientations. Approximately oriented candidate boxes are generated to the query image attributes (24) 15 suitable for the application of the candidate interest region finding process independent of orientation (26) A single-shot object detection result is obtained in this way (44). In this way, query-support in the images (11) (12) oriented objects single shot object detection (44) with accurate This ensures that the object is located and positioned accordingly. Additionally, single-shot object detection is enabled. In the method (10), due to the small number of positive support images (12), a small number Despite the number of positive examples it might encounter, orientation information differs from other attributes. It provides a more accurate learned solution through its production. In the invention, High quality labeling is achieved with minimum hardware load and labeling cost through the use of the developed method. With this accuracy, it is possible to detect object regions on the image. In the method, the approach uses the input (query) image and the support images (11) (12) in the same convolutional Feature extraction is performed by passing through the network (18) and in the feature extraction step, the traditional 25 Unlike other methods, a neural network that is not affected by orientation and features are directional. This ensures that it is learned and produced. The invention is not affected by orientation. It starts with finding the attributes; at this stage, the query and support images are shared. It becomes input to the convolutional network model, and in the next stages, the query image (11) is processed. Steps for extracting directed candidate boxes and candidate regions of interest using attributes 30 Although the orientation of the supporting image is different because it is applied, the semantic Similarities are better revealed. Query image (11) in the first stage The feature extraction network is input with the support image (12). In this way, training Training is provided with fewer parameters at this stage. The invention is independent of any orientation. Applying the candidate region of interest finding process (26) supporting images attribute (28) Applying a traditional candidate region of interest finding procedure to the tensor (27) This ensures that the feature extractor operates at the instance level without being affected by orientation. Attributes are extracted. In the invention, 5 of the traditional candidate region of interest finding process are used. Following the implementation (28), to generate features at the sample level with orientation-independent candidate interest region (30) and traditional candidate interest region procedures (32) It is possible to obtain a query candidate region of interest result (34). In this way, the query By obtaining sample-level results from the images (11), the merging step is taken. It can be transferred. In addition, the invention reduces the traditional candidate region of interest finding process by 10 Following the implementation (28), to generate features at the sample level with orientation-independent candidate interest region (30) and traditional candidate interest region procedures (32) It is ensured that a support candidate region of interest result is obtained (36). In this way, support By obtaining sample-level results from the images (12), the merging step is taken. It can be transferred. Then the query candidate interest region result and supporting candidate interest 15 After obtaining the result of the region (34) (36), a query candidate region of interest and a support The results corresponding to the candidate's area of ​​interest are combined (38). In this way The result obtained by combining the query candidate interest region result with the support candidate interest region result. Classification and regression can be performed on the obtained data. Here, trained... Classification from data is achieved using fewer processing steps by combining data, and 20 Regression is achieved. Then, the example is provided by combining the results in the invention. Classification and regression with object detection process to attributes (40) at the level (42) steps are performed. In this way, features at the sample level are obtained. It is. Approximately, from the data obtained by performing classification and regression (42), only one It is ensured that the result of the projectile object detection (44) is obtained. In this way, example 25 Object detection is performed using features obtained at this level. Figure 2 shows a study of a single-shot object detection method, which is the subject of the invention. The flowchart of the scenario is shown. The subject of the invention is the detection of a single-shot projectile. in the method (10), query and support images (11) (12) orientation unaffected feature 30 Feature tensors are obtained by passing them through the extraction network (Siamese convolutional network) (13). is done. The tensor of the query image (11) is first used in candidate box finding templates (20) oriented candidate boxes (22) are generated from the query image attributes (24). 6 There is a candidate interest region (30) independent of orientation. Support images (12) are: feature extraction network (13) of support images that are not affected by orientation The traditional candidate region of interest finding process is applied with the tensor (27) (28). By combining the results, (38) feature generation at the sample level is ensured. In this way, object detection from attributes at the instance level identifies objects belonging to the relevant classes. 5 Location and class information is generated. The invention involves the detection of a single-shot projectile. In the scenario shown in Figure 2 in the method (10), the oriented candidate box extractor This is demonstrated. The invention applies a traditional feature extraction process (Resnet). by replacing each layer with the convolutional mesh mentioned in the candidate box extractor (e2cnn). The method created (ReResnet[2]) is used. The subject of the invention is the detection of a single-shot object. The orientation-independent feature is extracted thanks to the method (10). Φ (orientation-independent feature) (independent convolutional network) where I is the input image and Tr is the orientation operation. If the mathematical equality Φ(𝑇𝑟Ι ) = 𝑇𝑟Φ(Ι ) satisfies, then the convolutional network orientation They are independent. The number of bands K for the features (𝐾, 𝑁, 𝑊, 𝐻) produced as a result of the convolutional network, N is the orientation number, H is the height, and W is the width. There are 15 orientations. an unaffected feature extractor network as a common network for support and query image. It is used. The resulting query image (11) features are used with the directed candidate box. An extractor is applied. Different methods are used during the candidate box extraction phase on the attributes. Candidate boxes of various sizes, aspect ratios, and orientations are being studied. Candidate boxes are categorized by left, top, height, width, and orientation as follows: 𝑥, 𝑦, 𝑤, ℎ, 𝜃 20 It includes the parameters. Candidate box outputs are in 6 different predetermined orientations, 3 by learning from candidate boxes produced in different sizes and 3 different aspect ratios It is being produced. Figure 3 shows a study of a single-shot object detection method given in Figure 2. The scenario shows the distribution of candidate boxes with different characteristics. The subject of the invention is... In the scenario shown in Figure 2 (10) of a single-shot object detection method, orientation 25 The feature extractor is shown at the unaffected sample level. This stage consists of two steps. It consists of the first stage oriented candidate obtained from the query images (11). The aim is to extract candidate areas of interest using boxes. Candidates independent of orientation. Method of extracting regions of interest: spatial and directional alignment (300) (302) It consists of sections. The orientations of the candidate boxes created in an orientational manner are 30 orientation-independent candidates to overcome the difficulty arising from being different. The regions need to be generated. In the spatial alignment step (302), 𝑥, 𝑦, 𝑤, ℎ, 𝜃 Parameters are generated. Directional alignment of the resulting misaligned outcome. Candidate regions with different orientations are generated with the number of orientations N for (302). Limited 7 Interpolation is applied to the intermediate values ​​of the number of orientations. In the second stage... The obtained orientation-independent candidate region of interest support for the query image (11) candidate interest generated by traditional candidate interest region extraction method for image (12) By combining these regions, a sample-level feature extractor is generated. Below are 5 orientations from (10) in a single-shot object detection method, which is the subject of the invention. The formula and explanations for the result of the unaffected candidate's area of ​​interest are given; 𝑓𝑅 = 𝐼𝑛𝑡(𝑆𝐶(𝑓𝑅 , 𝑟), 𝜃), 𝑟 = ⌊ 𝜃𝑁 2π⁄ ⌋ 𝐼𝑛𝑡 = 𝑖𝑛𝑡𝑒𝑟𝑝𝑜𝑙𝑎𝑠𝑦𝑜𝑛 𝑆𝐶 = 𝑦ö𝑛𝑒𝑙𝑖𝑚 𝑑𝑒ğ𝑖ş𝑡𝑖𝑟𝑚𝑒 𝑓𝑅 = 𝑦ö𝑛𝑒𝑙𝑖𝑚𝑑𝑒𝑛 𝑒𝑡𝑘𝑖𝑙𝑒𝑛𝑚𝑒𝑦𝑒𝑛 𝑎𝑑𝑎𝑦 𝑖𝑙𝑔𝑖 𝑏ö𝑙𝑔𝑒𝑠𝑖 𝑠𝑜𝑛𝑢𝑐𝑢 10 The subject of the invention is a single-shot object detection method in the scenario shown in Figure 2 (10). Classification and regression process (42) is shown. From the features at the sample level Accuracy maps of bounding box locations of objects with estimated bounding boxes A comparison is made. The error resulting from the difference is identified using error functions. 15 This ensures that the error is minimized. Separate error functions are used for classification and regression. is used. For this purpose, any classification and regression method available in the literature is used. The error function becomes available. Ultimately, the single-shot object detection result. object location information and class scores are obtained by obtaining (44). 25 8 Reference Numbers Given in the Figure Single-shot object detection method 11 Interrogation images 12 Support images 13 Siamese convolutional network 14 Input provided 16 Feature extraction network processing 18. Providing the feature tensor. 19 Query image tensor Candidate box templates 22. Implementing the directional candidate box removal process 24 Query image attributes 26. The process of finding candidate interest areas independent of orientation. 27. Attribute tensor of support images 28. Implementing the traditional candidate interest area identification process. Candidates of interest regardless of orientation 300 Spatial alignment 302 Orientation alignment 32 Traditional candidate interest area procedures 34. Obtaining the result of the query candidate's region of interest. 36 Obtaining the result of the candidate interest area for support. 38. Combining the results 40 Attributes at the Sample Level 42. Classification and regression analysis 44 Obtaining the result of single-shot object detection.

Claims

9 REQUESTS 1. Query and support images (11) (12) are not affected by orientation and feature (14) Input to a common Siamese convolutional network (13) that enables extraction, attribute (16) query and support images (11) (12) are separately matched by passing through the extraction network. A single-shot object detection 5 which includes providing the feature tensor in the future (18). The method is a query that is suitable for providing the feature tensor (18). The image tensor (19) was processed with candidate box templates (20), and the candidate boxes Applying the candidate box extraction process oriented toward its production (22) and approximately oriented candidate boxes are generated from an orientation to the query image attributes (24) In order to be suitable for the application of the independent candidate interest area finding process (26) 10 It is characterized by obtaining a single-shot object detection result (44).

2. A single-shot object detection method compliant with Claim 1, characterized by its ability to support images. Applying a traditional candidate region of interest finding operation to the feature tensor (27) (28) is the inclusion of the process step. 15 3. A single-shot object detection method that complies with Claim 2, and whose characteristic is; at the sample level. independent candidate region of interest (30) and orientation toward generating attributes traditional candidate region of interest operations (32) and obtaining a query candidate region of interest result (34) is the process step. 20 4. A single-shot object detection method that complies with Claim 2, and its characteristic is that it is at the sample level. independent candidate region of interest (30) and orientation toward generating attributes traditional candidate interest region operations (32) and obtaining a support candidate interest region result (36) is the process step. 25 5. A one-shot object detection method that conforms to claims 3-4, characterized by its ability to identify a query candidate interest. Combining the results corresponding to the region and a support candidate region of interest (38) It includes the process step.

6. A single-shot object detection method conforming to claim 5, characterized by its ability to combine results. Classification and object detection process with the provided sample level attributes (40) Performing regression involves (42) steps.

7. Is there a single-shot object detection method that meets any of the previous requirements? 5 The feature is that a single-shot object detection result is obtained (44) classification and This is obtained from the data obtained by performing regression (42).