A fine identification method for multi-class aircrafts in high-resolution remote sensing scenes

Abstract

The application discloses a fine identification method for multi-class aircrafts in a high-resolution remote sensing scene, relates to the fields of image processing and deep learning, and comprises the following steps: high-resolution remote sensing image data is acquired, and an aircraft target data set is prepared; a high-resolution remote sensing image to be detected is inputted, and the image is preprocessed; a fine identification model is trained; the preprocessed high-resolution remote sensing image is sent into the fine identification model for detection, and a detection result is obtained; and the detection result is spliced and mapped to an original image. The application can relieve the problem that a conventional target detection method has difficulty in detecting small targets in a high-resolution remote sensing image, has finer classification and positioning capabilities, has more advantages and generalization capabilities in a remote sensing scene, and has stronger detection performance.

Description

Technical Field

[0001] This invention relates to the fields of image processing and deep learning, and in particular to a method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes. Background Technology

[0002] In recent years, with the continuous improvement of remote sensing technology, the number and quality of remote sensing images of the Earth have been increasing. Sufficient remote sensing datasets have provided data support for deep learning-based remote sensing image research, driving the development of remote sensing image processing. An important topic in image processing research is detecting specific targets within images, a process known as target detection.

[0003] Unlike target detection in conventional scenarios, multi-class aircraft target detection in remote sensing scenarios is extremely difficult, mainly due to two reasons. Firstly, the nature of the targets themselves: the detection task involves multiple categories of aircraft, and the differences between some categories are not significant, with many similar features making them difficult to distinguish in remote sensing scenarios. Secondly, high-resolution remote sensing images cover a large area; even aircraft that are large targets in natural scenes only occupy tens of pixels in remote sensing images, a very small percentage in high-resolution images with tens of millions of pixels. The large remote sensing dataset DOTA defines targets with 10-50 pixels as small targets, so aircraft occupying only tens of pixels are also considered small targets. Thirdly, the nature of remote sensing image capture: satellites are thousands of miles away from the ground, and cloud cover can easily obscure targets that are easily observed on the ground, making them difficult to detect. Furthermore, remote sensing images are captured by satellites around the clock, and exposure issues, such as nighttime conditions, severely interfere with image quality, further increasing the difficulty of detection.

[0004] YOLOv5 is a single-stage target detection algorithm known for its fast detection speed, high accuracy, numerous model interfaces, ease of expansion, and suitability for industrial deployment, making it widely used in natural scenes. However, due to the limitations of remote sensing scenarios, the YOLOv5 model, which performs excellently in natural scenes, struggles to achieve good detection results in remote sensing environments. Therefore, this invention improves upon YOLOv5, proposing a refined identification method for multiple aircraft categories in high-resolution remote sensing scenes based on optimized YOLOv5, which addresses the aforementioned issues. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] In view of the problems existing in the above and / or existing methods for fine identification of multiple types of aircraft in high-resolution remote sensing scenarios, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to provide a method for fine identification of multiple types of aircraft in high-resolution remote sensing scenarios.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for fine recognition of multiple types of aircraft in high-resolution remote sensing scenes, comprising: acquiring high-resolution remote sensing image data and creating an aircraft target dataset; inputting a high-resolution remote sensing image to be detected and preprocessing the image; training a fine recognition model; feeding the preprocessed high-resolution remote sensing image into the fine recognition model for detection and obtaining detection results; and stitching the detection results together and mapping them onto the original image.

[0009] As a preferred embodiment of the fine identification method for multiple types of aircraft in high-resolution remote sensing scenarios described in this invention, the aircraft target dataset includes a training set, a validation set, and a test set, wherein the training set and the validation set are labeled with the location and category information of all aircraft targets.

[0010] As a preferred embodiment of the fine identification method for multiple types of aircraft in high-resolution remote sensing scenarios described in this invention, the preprocessing includes cropping the high-resolution remote sensing image using an overlapping sliding window, wherein the sliding window has a preset size of 640×640 and a preset overlap interval of 100.

[0011] As a preferred embodiment of the high-resolution remote sensing scene multi-category aircraft fine recognition method described in this invention, the training of the fine recognition model includes, during training, inputting training set images as model feature learning samples and inputting validation set images as model evaluation and validation samples. During evaluation, the validation index is the average accuracy of each category of targets. The higher the accuracy, the better the weight file trained by the model. The best-performing weight file is selected and sent to the test set for testing.

[0012] As a preferred embodiment of the high-resolution remote sensing scene multi-type aircraft fine recognition method described in this invention, the fine recognition model includes: a backbone network, a neck feature fusion network, and a detection module; the backbone network includes, in sequence, a convolutional component 1, a convolutional component 2, a CSP1_1 component, a convolutional component 3, a CSP1_3 component, a convolutional component 4, an SPPCSPC module, and a deformable attention module; the feature fusion network includes a CSP2_1 component 1 connected to the deformable attention module, a strip attention module 1, an upsampling module 1, and a first stitching module connected to the CSP1_3 component, the obtained stitching result 1 is input in two directions through the sequentially arranged CSP2_1 component 2, one side is input in two directions through the sequentially arranged CSP2_1 component 2. Component 2, the long strip attention module 2, the upsampling module 2, and the second stitching module are connected to the CSP1_1 component. The resulting stitching result 2 is processed by the CSP2_1 component 3, which is set up in sequence, and input in two directions. On one hand, it is processed by the CSP2_1 component 3, the long strip attention module 3, the convolution component 5, and the third stitching module is connected to the CSP2_1 component 2. On the other hand, it is output to the decoupling head 1. The resulting stitching result 3 is processed by the CSP2_1 component 4, which is set up in sequence. On one hand, it is output to the decoupling head 2. On the other hand, it is processed by the long strip attention module 4, the convolution component 6, and the fourth stitching module is connected to the CSP2_1 component 1. The resulting stitching result 4 is output to the decoupling head 3 after passing through the CSP2_1 component 5. The detection module includes a decoupling detection head and a weighted bounding box fusion module.

[0013] As a preferred embodiment of the high-resolution remote sensing scene multi-class aircraft fine recognition method described in this invention, the SPPCSPC module is derived from YOLOv7, with more feature extraction layers and two parallel branches. The first branch uses three consecutive convolutional components, then connects four layers of max pooling with different windows, and finally connects two convolutional components. The second branch uses a single-layer convolutional component as a simple short-circuit connection, and finally the feature maps output by the two branches are fused.

[0014] As a preferred embodiment of the fine identification method for multiple types of aircraft in high-resolution remote sensing scenarios described in this invention, the deformable attention module learns multiple sets of deformable points through an offset network, then uses bilinear interpolation for feature sampling to obtain deformable keys and deformable values, and combines the relative position offset calculated from the deformable points with the data to input into a multi-head attention mechanism to obtain global information.

[0015] As a preferred embodiment of the high-resolution remote sensing scene fine identification method for multiple types of aircraft described in this invention, the strip attention module achieves the following effects: inputting feature maps and sending them to two branches in parallel; connecting modules in the attention mechanism to enhance the local spatial information of the target.

[0016] As a preferred embodiment of the high-resolution remote sensing scene multi-type aircraft fine identification method described in this invention, the convolution component is composed of 3×3 convolution, batch normalization and activation function in sequence.

[0017] As a preferred embodiment of the high-resolution remote sensing scene multi-category aircraft fine recognition method described in this invention, the decoupled detection head includes: firstly adjusting the number of channels using a 1×1 convolution, then using two parallel branches, each branch containing a 3×3 convolution, and adjusting the channels again using a 1×1 convolution. The first branch is responsible for target classification, and the second branch is further divided into two branches, one responsible for regression and the other responsible for determining confidence. The weighted box fusion retains the best candidate box and two second-best candidate boxes for each target, and uses the confidence scores corresponding to the three candidate boxes as weights to adjust the position of the target box.

[0018] The beneficial effect of this invention is that it alleviates the difficulty of detecting small targets in high-resolution remote sensing images using conventional target detection methods. By compressing the network structure, the network model obtains more shallow feature information, and SPPCSPC and deformable attention are used to construct global relationships. For aircraft targets, a striped attention mechanism is designed to improve the ability to mine local features of aircraft targets, and a decoupled head detection and weighted bounding box fusion scheme is used to give the model more refined classification and localization capabilities. Compared with target detection schemes in conventional scenarios, this invention has greater advantages in remote sensing scenarios. Compared with the baseline model, the mAP index of this model is significantly improved, reflecting the effectiveness of the model design, demonstrating greater generalization ability and stronger detection performance in remote sensing scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0020] Figure 1 This is a diagram illustrating the fine recognition architecture of a method for fine recognition of multiple aircraft types in high-resolution remote sensing scenarios.

[0021] Figure 2 A comparison diagram of the backbone networks for fine identification methods of multiple aircraft types in high-resolution remote sensing scenarios.

[0022] Figure 3 This is a bar attention map for a refined identification method of multiple aircraft types in high-resolution remote sensing scenes.

[0023] Figure 4 A visualization comparing detection methods for fine-grained identification of multiple aircraft types in high-resolution remote sensing scenarios.

[0024] Figure 5 This is a visualization of the original image mapping for a sophisticated method of identifying multiple types of aircraft in high-resolution remote sensing scenes.

[0025] Figure 6 A data augmentation map for a refined identification method of multiple aircraft types in high-resolution remote sensing scenarios. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0028] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0029] Example 1

[0030] Reference Figures 1-5 This is the first embodiment of the present invention, which provides a method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes includes:

[0031] S1: Acquire high-resolution remote sensing image data and create an aircraft target dataset. Specifically, the dataset used is the Rsaicp visible light aircraft image dataset, which contains 11 types of aircraft targets: A (bombers), B (helicopters), C (transport aircraft), D (fighter jets), E (tanker aircraft), F (anti-submarine patrol aircraft), G (early warning aircraft), H (unmanned aerial vehicles), I (Osprey tiltrotor aircraft), J (passenger aircraft), and K (other aircraft). The dataset is divided into training set, validation set, and test set.

[0032] S2: Input the high-resolution remote sensing image to be detected and preprocess it. Specifically, preprocessing is performed before the high-resolution remote sensing image is input. The high-resolution remote sensing image is cropped into smaller images of the same size, each image being 640×640 pixels, with an overlap of 100 pixels. The cropping scale is adjusted according to a size factor, with a default size ratio factor of 1. It should be noted that during the cropping process, the positions of the target labels corresponding to the original image are also adjusted and updated accordingly.

[0033] S3: Training the Refined Recognition Model. Specifically, during the training and testing phases, the input images are resized to 640×640 before being fed into the network. Since all small images are 640×640 after cropping, no adjustment is made by default. The network uses SGD and Momentum as optimizers, with a batch size of 32, 100 epochs, and the rest being YOLOv5 default parameters. The model is trained on the training set in multiple batches, and its performance is validated using a validation set. The validation metric is the average accuracy across all target categories; higher accuracy indicates a better-performing weight file. The best-performing weight file is then tested on the test set.

[0034] S4. The preprocessed high-resolution remote sensing image is fed into the fine recognition model for detection, and the detection result is obtained.

[0035] The refined recognition model consists of a backbone network, a feature fusion network, and a detection module. Specifically, the backbone network comprises a compressed CSP Display, SPPC-SPC, and a deformable attention mechanism. While compressing the backbone network, the output feature layer is shifted forward, resulting in output sizes of 160×160×255, 80×80×255, and 40×40×255. SPPC-SPC is used to further enhance feature extraction, and a deformable attention mechanism is employed to construct global relationships at a lower cost.

[0036] Specifically, the backbone network includes, in sequence, convolutional component 1, convolutional component 2, CSP1_1 component, convolutional component 3, CSP1_3 component, convolutional component 4, SPPCSPC module, and deformable attention module;

[0037] During feature extraction, three layers of feature maps at different scales are obtained and fed into the feature fusion part of the neck region. In the feature fusion part, a strip attention mechanism is designed to capture the local features of the aircraft target. Its idea originates from Strip Pooling, and combined with the detection scenario, it enhances the local features of the aircraft target. This attention mechanism is not only applicable to aircraft, but can also enhance the features of rectangular targets using appropriate strip convolution kernels. In the feature fusion module, a strip attention mechanism is set after each CSP2_1 module. CSP2_1 is the main feature extraction component of YOLOv5. The feature map output from the CSP2_1 component contains rich local information, thus enhancing the potential features of the aircraft target from this rich information and suppressing useless background information.

[0038] Specifically, the feature fusion network includes a CSP2_1 component 1 connected to a deformable attention module, a strip attention module 1, an upsampling module 1, a first stitching module connected to a CSP1_3 component, and the resulting stitching result 1 is fed into a CSP2_1 component 2 in two directions. On one hand, it is fed into a CSP2_1 component 2, a strip attention module 2, an upsampling module 2, a second stitching module connected to a CSP1_1 component, and the resulting stitching result 2 is fed into a CSP2_1 component 3 in two directions. On one hand, it is fed into a CSP2_1 component 3, a strip attention module 3, a convolutional component 5, a third stitching module connected to a CSP2_1 component 2, and the result is output to a decoupling head 1. The resulting stitching result 3 is fed into a CSP2_1 component 4 in two directions. On one hand, it is output to a decoupling head 2. On the other hand, it is fed into a strip attention module 4, a convolutional component 6, a fourth stitching module connected to a CSP2_1 component 1, and the resulting stitching result 4 is output to a decoupling head 3 via a CSP2_1 component 5.

[0039] Furthermore, the strip attention mechanism is designed based on the appearance characteristics of aircraft targets, enabling the construction of local relationships within a single aircraft target. Specifically, a feature map of size H×W×C (H, W, and C are the length, width, and dimension of the feature map, respectively) is fed into two branches in parallel after input. The first branch uses two layers of parallel max pooling with windows of 5 and 9, concatenating the feature maps before feeding them into three layers of parallel convolutions with windows of 7×7, 7×1, and 1×7. After simple concatenation, a fully connected layer and a sigmoid activation function are used to generate attention weight maps. The second branch uses a simple convolutional component, multiplying the output of the first branch, and then connecting it to the spatial attention module in the convolutional attention mechanism to enhance the local spatial information of the target.

[0040] The convolutional component consists of a 3×3 convolution, batch normalization, and activation function.

[0041] The detection module consists of a decoupled detection head and weighted bounding box fusion. Within the decoupled detection head, a 1×1 convolution is first used to adjust the number of channels. Then, two parallel branches are used, each containing a 3×3 convolution. After adjusting the channels again with a 1×1 convolution, the first branch handles target classification, and the second branch further splits into two branches: one for regression and the other for determining confidence. Weighted bounding box fusion retains the best candidate box and two second-best candidate boxes for each target. The position of the target box is weighted using the confidence scores of these three candidate boxes as weights, aiming to minimize missed detections. The detection head follows the YOLOv5 multi-scale detection head approach, using a three-layer head to capture as many potential real targets as possible and provide a sufficient number of candidate boxes for weighted bounding box fusion. This method does not use the coupled detection head of YOLOv5 but decouples it, no longer sharing parameters between classification and regression tasks. Although this increases model inference time, it achieves more accurate detection results. The candidate boxes obtained by the three-layer detection head are sent to the weighted box fusion module, sorted from high to low according to the target confidence, retaining the best candidate box and two second-best candidate boxes near each target, and then calculating the weighted average of the three according to the confidence to obtain the final detection result, which alleviates the problem of inaccurate target localization in remote sensing scenarios.

[0042] S6. The detection results are stitched together and mapped onto the original image. Specifically, after detection, the target location and category information detected in the small image are stitched onto the corresponding location in the high-resolution large image. For targets that are repeatedly detected in the overlapping areas after cropping, non-maximum suppression is used for simple filtering. The final result is mapped onto the original image to visualize the detection.

[0043] As shown above, the method provided by this invention can alleviate the difficulty of detecting small targets in high-resolution remote sensing images using conventional target detection methods. By compressing the network structure, the network model obtains more shallow feature information, and SPPCSPC and deformable attention are used to construct global relationships. For aircraft targets, a striped attention mechanism is designed to improve the ability to mine local features of aircraft targets, and a decoupled head detection and weighted bounding box fusion scheme is used to give the model more refined classification and localization capabilities. Compared with target detection schemes in conventional scenarios, this invention has greater advantages in remote sensing scenarios. Compared with the baseline model, the mAP index of this model is significantly improved, reflecting the effectiveness of the model design, demonstrating greater generalization ability and stronger detection performance in remote sensing scenarios.

[0044] Example 2

[0045] Reference Figure 6This is the second embodiment of the present invention. To verify the effectiveness of Example 1, the present invention conducted an experiment using the Rsaicp visible light aircraft image dataset as an example. Due to the small number of G-class aircraft, the detection effect was found to be poor after the experiment. Therefore, data augmentation was performed on G-class aircraft. By filtering out the images containing G-class aircraft and performing schemes such as exposure, rotation, and cropping the target to other background images, the number of G-class targets was increased. The figure shows that after data augmentation, the number of G-class targets is 5 times that of the original. Since there are some other types of aircraft in the images containing G-class aircraft, the number of other types of aircraft also increased slightly.

[0046] The evaluation metrics used in the experiment were the average precision (AP) for a single class and the average precision (mAP) for all targets. Higher AP and mAP scores indicate better detection performance. Table 1 shows the detection comparison results of our method with other methods on the Rsaicp visible light aircraft image dataset; Table 2 shows the ablation experiment results of our method on the Rsaicp visible light aircraft image dataset.

[0047] Table 1. Comparison of detection results between this method and other methods.

[0048]

[0049] The model presented in this study improved the mAP metric by 12% compared to the baseline model YOLOv5. After augmenting the target data for class G, the mAP metric improved by another 4%.

[0050] Table 2 Ablation Experiment Results of This Method

[0051]

[0052] The table clearly demonstrates that the model of this invention has higher detection accuracy and can better perform its detection function. The model of this method improves the mAP index by 12% compared to the baseline model at AP(IoU 0.5) and by 17% compared to the baseline model at AP(IoU 0.7).

[0053] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for fine identification of multiple aircraft types in high-resolution remote sensing scenes, characterized in that: Acquire high-resolution remote sensing image data and create an aircraft target dataset; Input a high-resolution remote sensing image to be detected and preprocess the image; Train a refined recognition model; The preprocessed high-resolution remote sensing image is fed into a fine recognition model for detection, and the detection result is obtained. The detection results are stitched together and mapped onto the original image; The refined recognition model includes a backbone network, a neck feature fusion network, and a detection module; The backbone network includes, in sequence, a convolutional component 1, a convolutional component 2, a CSP1_1 component, a convolutional component 3, a CSP1_3 component, a convolutional component 4, an SPPCSPC module, and a deformable attention module; The neck feature fusion network includes a CSP2_1 component 1 connected to the deformable attention module, a strip attention module 1, an upsampling module 1, and a first stitching module connected to the CSP1_3 component. The resulting stitching result 1 is processed by the CSP2_1 component 2, the strip attention module 2, the upsampling module 2, and the second stitching module connected to the CSP1_1 component. The resulting stitching result 2 is processed by the CSP2_1 component 3, and input in two directions. On one hand, it is processed by the strip attention module 3, the convolutional component 5, and the third stitching module connected to the CSP2_1 component 2. The resulting stitching result 3 is processed by the CSP2_1 component 4, and output to the decoupling head 2 on one hand. On the other hand, it is processed by the strip attention module 4, the convolutional component 6, and the fourth stitching module connected to the CSP2_1 component 1. The resulting stitching result 4 is output to the decoupling head 3 via the CSP2_1 component 5, and output to the decoupling head 1 on the other hand. The detection module includes a decoupled detection head and a weighted bounding box fusion module; The elongated attention module is an attention mechanism used to construct local relationships for a single aircraft target. The specific steps are as follows: The feature map of size H×W×C is fed into two branches in parallel after being input. The first branch uses two layers of parallel max pooling with windows of 5 and 9. After the feature maps are concatenated, they are fed into three layers of parallel convolution with convolution windows of 7×7, 7×1, and 1×7. After simple concatenation, a fully connected layer and a sigmoid activation function are used to generate an attention weight map, where H, W, and C are the length, width, and dimension of the feature map, respectively. The second branch uses a convolutional component, and the output is multiplied by the first branch. It is then connected to the spatial attention module in the convolutional attention mechanism to enhance the local spatial information of the target. The convolutional component consists of a 3×3 convolution, batch normalization, and activation function.

2. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 1, characterized in that: The aircraft target dataset includes a training set, a validation set, and a test set, wherein the training set and the validation set are labeled with the location and category information of all aircraft targets.

3. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 1 or 2, characterized in that: The preprocessing includes cropping the high-resolution remote sensing image using an overlapping sliding window, wherein the sliding window has a preset size of 640×640 and a preset overlap interval of 100.

4. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 3, characterized in that: The trained fine-grained recognition model includes, During training, the images in the training set are used as model feature learning samples, and the images in the validation set are used as model evaluation and validation samples. During model evaluation, the validation metric is the average accuracy of each category of targets. The higher the accuracy, the better the weight file trained by the model. The best-performing weight file is then selected for testing on the test set.

5. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 4, characterized in that: The SPPCSPC module uses two parallel branches. The first branch uses three consecutive convolutional components, then connects four layers of max pooling with different windows, and finally connects two convolutional components. The second branch uses a single convolutional component as a simple short-circuit connection. Finally, the feature maps output by the two branches are fused.

6. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 5, characterized in that: The deformable attention module learns multiple sets of deformation points through an offset network, then uses bilinear interpolation for feature sampling to obtain deformation keys and deformation values. Combined with the relative position offset calculated from the deformation points, it is fed into a multi-head attention mechanism to obtain global information.

7. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 6, characterized in that: The convolutional component consists of a 3×3 convolution, batch normalization, and an activation function.

8. The method for fine identification of multiple types of aircraft in high-resolution remote sensing scenes as described in claim 7, characterized in that: The decoupled detection head first uses a 1×1 convolution to adjust the number of channels, then uses two parallel branches, each branch containing a 3×3 convolution, and then uses a 1×1 convolution to adjust the channels again. The first branch is responsible for target classification, and the second branch is further divided into two branches, one responsible for regression and the other responsible for determining confidence. The weighted box fusion method retains the best candidate box and two second-best candidate boxes for each target, and uses the confidence scores of the three candidate boxes as weights to adjust the position of the target box.

Images